Skip to content

Toskar / What's new

What's new

Every release of Toskar Core and what it changed, from the project's changelog. Toskar Pro is built on Core, so these changes reach it too.

1.14.1

November 19, 2026

Downloads and release notes ↗Docs for 1.14.1 →

Added

  • Each answer now records which model wrote it and which computer it ran on, so the phone app can say "Llama 3.2 1B on Studio" under it (contract 1.109, toskar-apps#447).
  • Let an automation act on its own: under Advanced, choose a Home Assistant device and what to do, such as turn on the porch light. A run may make that exact call without asking, even when a webhook, device, or camera started it; any other call that changes things is still skipped.
  • Connect your own ComfyUI: an Admin adds its address on Tools › Images and video › ComfyUI and saves workflows exported in API format, such as an upscale or pictures in a set style. Chats run them with the ComfyUI Workflow tool, and what they make is attached to the answer. Toskar finds each workflow's prompt, seed, picture, and size inputs, and never installs anything into ComfyUI.
  • An automation that speaks at the door can now talk with the visitor. Choose Talk with the visitor and, after its first words, Toskar listens through a Reolink doorbell's microphone, answers, and finds out who came and why, taking turns and never saying whether anyone is home. When the visit ends, you're notified with a short summary and what was said. Hearing the doorbell uses reolink-cli's live preview and ffmpeg.
  • An automation can speak through a Reolink doorbell: choose Say the answer at the door, and each run's answer is read aloud by this computer's voice through the doorbell's speaker. With the doorbell's picture, it says nothing to someone just passing by, and quiet hours, such as 10 PM to 7 AM, keep it to a notification. It uses Reolink's reolink-cli, with the doorbell added there and its name given as the connection's Speaker.
  • Model downloads can go through a download cache on your own network: set TOSKAR_DOWNLOAD_MIRROR to its address, and Toskar asks it before Hugging Face, falling back to Hugging Face when the cache fails or its copy doesn't match the model's checksum.
  • Look up a flight's status: connect Flight status (AeroDataBox) under Tools › Connected services with your own key, then ask "What gate is AA3595 leaving from?" for the gate, terminal, scheduled and expected times, delay, and baggage belt.
  • Profiles can talk with visitors: turn on Talks with visitors on the Topics tab, or start from the new Front door template. Toskar's rules for strangers come first: it says it's an automated assistant, never says whether anyone is home or who lives there, won't open a door or turn off an alarm for anyone, and only uses tools that change nothing. None of your style, location, memories, or computers reach its chats. A new quality set checks these guardrails in several languages (contract 1.108).
  • More natural voices in 15 languages: an Admin can install Kitten TTS 2 mini from Settings, under Voice, and it then reads the languages Kokoro doesn't speak, such as German, Korean, Russian, Dutch, and Hindi, instead of Piper. Voice also lets an Admin choose the voice for each language, with a preview, for voice mode, answers read aloud on the phone and in the car, and the doorbell. An automation that speaks at the door can have a voice of its own. Kitten's voices are Powered by KittenML, under the Apache-2.0 license.
  • Live voice mode's connection (#795, contract 1.100): a WebSocket at /api/v1/voice/realtime that speaks OpenAI's Realtime protocol. Toskar hears when you stop talking, answers through your chat, speaks the answer a sentence at a time, and stops when you talk over it. Each answer reports how long you waited. See docs/voice.md.
  • Every model in the catalog says what it does: a category, such as chat and coding, search and retrieval, or creative, the tasks it handles, and what it takes in and gives back. The API lists them, ready for browsing the store by category.
  • A model whose license has terms to accept, such as Gemma or Llama, shows them before it's installed from Models: read them, then Accept and install (#593). Toskar keeps a record of who accepted which license and when, for the Admin (contract 1.103: accept_license on install, license.accepted_at, and GET /api/v1/models/license-acceptances).
  • The model store says what each model's license allows (#593): Free, Free with conditions, or License unknown, and whether a business may use it. For business use shows only models whose license allows that. Every model in Toskar's catalog has its license checked by hand (contract 1.102: license on catalog models, commercial and license_kind on store listings, commercial on picture and clip models).
  • Models › Installed says when a newer model replaces one you have, such as Qwen 3.5 4B for Qwen 2.5 3B, with how it fits and whether its license differs, and installs it beside the old one. Admins get one notification per new generation, which Settings › Notifications can turn off.
  • Type your town under Settings → Personalization → Your location, and ask "Coffee near me?", "What's the weather?", or "Directions to the airport" without naming a place. It's used only for questions like these, and the answer says when it used it.
  • A pending Access from anywhere checkout can be cancelled from API Access.
  • A Skip to content link, the first thing on every page for keyboard users.
  • A subscription tied to another computer can be moved to this one from API Access.
  • Automations warn when the chosen model can't use the tools they need.
  • Setup offers "Set up later" when another page was opened, and goes there.
  • Connect a Reolink camera or doorbell straight through the camera on your network, with no Home Assistant or Reolink hub software: under Connected services, give its address, a user, and a password, or for a doorbell paired to a Home Hub, the hub's address and the doorbell's channel. An automation can look at it as it runs, or start when it sees a person, a vehicle, an animal, or motion, or when someone rings the doorbell.
  • POST /api/v1/speech/say (#795, contract 1.101) reads a sentence or a few aloud in live voice mode's voice, Kokoro or Piper, and answers the WAV file without keeping it. The phone, the car, and the TV use it to speak answers in the computer's natural voice. Paired devices may use it, with the voice capability.
  • Browse the model store by category: chat and coding, models that see pictures, and picture and clip models, which you can now set up from Models as well as from Tools.
  • Make a clip longer: ask "Continue the clip: now a surfer paddles out" or "Make the clip longer", and Toskar makes the next part from the clip's last frame and joins the two into one clip.
  • Make a longer video: ask for "a 1-minute video of a day at the beach", and Toskar plans it as shots of up to 5 seconds, shows you the shots and about how long they take before anything is made, then makes each shot from the last frame of the one before and joins them into one video.
  • Voice mode in the desktop app (#795). The app's window opens live voice's WebSocket on the daemon itself, since the shell's proxy can't carry one, at the address the shell gives. It needs the Toskar app from 1.14 or later.
  • Voice mode in the web app (#795): choose Talk beside Send to talk with Toskar out loud. It hears when you stop, answers in a voice as it writes, and stops when you talk over it. The user guide's Talk with Toskar out loud page explains it.
  • Live voice's WebSocket takes the API key as an openai-insecure-api-key.<key> subprotocol, as OpenAI's browser clients send it.

Changed

  • A reasoning model can keep its thinking between tool calls in one answer, instead of working out its plan again after each call. It's off until the quality set shows it helps; TOSKAR_KEEP_REASONING=field or inline turns it on to compare.
  • Access from anywhere leads with its switch: the plans and the router setting wait until it's on.
  • Adding a nearby computer has one button; approving with a code is behind a link on its card.
  • The sidebar's status counts other computers, says "Just this computer" on its own, and shows Knowledge only to admins.
  • Tools shows a short description for people; the model keeps its own.
  • The sidebar's Placement and Knowledge rows show real status, or stay out.
  • Access from anywhere keeps its router settings out of the way while off, and says which port opens before it's turned on.
  • The deploy list on Computers puts what fits first and leaves out the test model, and "Manage" is now "Details".
  • The context meter beside Send shows its percentage.
  • Speaking and listening through a Reolink doorbell needs nothing more than reolink-cli installed on this computer. Toskar now reaches the camera with the connection's own address, channel and login, and runs reolink-cli's gateway at 127.0.0.1:9000 itself, so there's no camera to add to reolink-cli and no Speaker name to type. A camera you added to reolink-cli yourself, or a gateway elsewhere, can still be named in the connection.
  • Voice settings now say what voice mode and the doorbell still need: a chat model to answer with, and a model that can see a camera's picture. Each names the one that fits this computer, with its size, and installs it with one press. An automation that looks at a camera offers a model that can see when you choose the camera, instead of failing when someone is at the door. Installing Kitten TTS 2 mini now also sets up listening, so voice mode's first conversation doesn't wait for downloads.
  • Answers start sooner when the clock has moved on since the last message: the time now comes at the end of the AI's instructions, so the rest is reused instead of read again. Live voice feels it most.
  • Live voice mode keeps Whisper and the voices loaded between turns, so a turn no longer waits for them to start (#795).
  • Live voice mode speaks in a more natural voice: Kokoro, for English, Spanish, French, Italian, and Portuguese. Its model downloads the first time; Piper reads until then, and reads the other languages.

Fixed

  • A doorbell that can't be answered is said plainly: when Toskar can't reach the camera's speaker, you're told so in your language, with no raw reolink-cli output and no retry that would greet an empty porch minutes later. Toskar restarts only the reolink-cli gateway at its own address, the door's notes and transcript are in your language, and the instant greeting covers 20 more languages, such as Dutch, Russian and Swedish (#1340).
  • API Access is clearer (#1266):
  • With a developer's relay, the subscription card names that relay instead of calling it Toskar's.
  • Phones and tablets counts only devices in use, and says how many keys haven't been used lately.
  • A device connected through a phone without a name of its own is listed as "Device via" that phone.
  • Finding a setup during onboarding says what it's checking while it works.
  • When an automation's run answers without opening the page its instructions name, as a model that can't use tools does, you get a notification saying so and suggesting a model that can (#1269).
  • The boot screen no longer waits for good on a health check that never answers: each check gives up after six seconds and is tried again, and Try again starts a new one (yeixio/toskar-apps#461).
  • The built-in profiles (General Assistant, Programming, Research, Custom) and "the AI news" in What left this computer now show in your language on the web and in the phone app, unless you renamed the profile (#1271).
  • The certificate a computer shows anyone who connects, from the internet or through the relay, no longer has its name. Certificates made before carry it, so Toskar makes a new one and switches once the paired phones have learned it, or after 30 days. Phones with an app from before see the usual "certificate changed" check once (#1263).
  • While Toskar starts on an answer, "Starting…" and "Loading the model" show right under your question instead of down by the message box, and the Latest button sits in the corner on wider windows so it doesn't cover the end of an answer (#1270).
  • Changing one setting of a connected service, such as a Reolink camera's Speaker, no longer asks for every other setting again ("Camera address is required"). The form keeps what's stored and saves only what you change.
  • A connection form's example values now read "e.g. front-door" rather than a bare "front-door", so they don't look like values already filled in.
  • The Mac and Windows apps' "already running" dialog no longer calls every other Toskar "Toskar Pro".
  • The Mac and Windows app offers guided setup on a new install, even when the window remembers setup from an earlier one.
  • Phones, tablets and TVs that didn't buy access from anywhere keep finding the computer away from home after its subscription renews: they get the newer device token from GET /api/v1/remote-access/device-token, from anywhere, instead of only at home (yeixio/toskar-apps#451, contract 1.99).
  • Talk with the visitor now hears someone speaking at a Reolink doorbell. The doorbell's microphone is quiet, and Toskar listened only for speech about twice as loud as a visitor's voice, so it greeted them and then ended the visit as if nobody had answered.
  • What Toskar says through a Reolink doorbell is louder: the voice is raised to near full volume before it's sent, since the doorbell's small speaker made it hard to hear. It also leaves out the low end the small speaker can't play and evens out quiet and loud syllables, so it sounds about 7 dB louder at the same peak.
  • The doorbell greets a visitor sooner: its voice, and its listening, are readied while Toskar looks at the camera and writes the greeting, rather than started after.
  • A visit's notification is plain sentences, without headings, lists, or notes about Toskar's own settings, and the doorbell no longer calls a visitor by a name they only mentioned, such as who sent a package.
  • After the doorbell says goodbye, it keeps listening for a minute in case the visitor has more to say, instead of ending the visit at once.
  • With Talk with the visitor on, the doorbell greets someone the moment it sees them ("Hello! How can I help you?", in the automation's language), instead of first looking at the picture and writing a greeting, which took about 20 seconds. Someone passing by hears the greeting too.
  • A Member's phone or key no longer reads where the computer keeps its files, the address it listens on, its node ID, or its address from outside in Settings.
  • Asking for the summary of a chat that doesn't exist, or compacting one, answers "not found" instead of a server error.
  • The recommended setup in onboarding no longer carries a "Mimir" badge, which read like a model or vendor name (#1251).
  • A router port that an older router kept open with no time limit is closed at the next start when Toskar crashed or was killed while it was open.
  • A key made in the "Allow access" dialog now lets you allow access.
  • Removing a model from a computer asks first.
  • Adding one nearby computer no longer shows every card as busy.
  • Models calls this computer by its Toskar name, not the system's hostname.
  • Page watch runs when the page changes, as its summary says.
  • Enter in the Knowledge path field connects it.
  • Tall onboarding steps scroll instead of cutting off the top.
  • Sizes read "4 B", not "4 byte".
  • Messages point to where settings are now, such as Tools → Connected services and Computers → Network & access.
  • When the relay refuses this computer, API Access says why and what to do, instead of showing a code, and a subscription moved or refunded elsewhere shows as ended.
  • Toskar re-registers with the relay as soon as it reconnects, and takes its record back when Access from anywhere is turned off.
  • Toskar doesn't sell a subscription to a computer another device's subscription already covers.
  • Ports mapped on the router are released when Toskar stops.
  • A question the local knowledge answers doesn't search the web.
  • A push that reaches some devices retries only the ones that missed it.
  • Routers that take only permanent UPnP mappings get one.
  • A checkout page opened before a restart is still found after it.
  • Computers: the waiting card goes once the other computer joins, can be stopped, and says when there's no answer; the latest error shows, can be dismissed, and names the computer; several removals show at once.
  • API Access says HTTPS only when it's on, and shows no defaults under a load error.
  • The Access from anywhere switch moves at once.
  • Device keys never used, or not for 90 days, are marked.
  • The Models Store's memory note no longer contradicts the fit.
  • No startup screen flash when the service is already up.
  • Settings choices show which one is picked, clearly in light mode too.
  • Setup offers keep their details readable on phones.
  • Connecting a Reolink camera takes its address as the Reolink app shows it, such as 192.168.1.50, and tries HTTPS then HTTP. An address that doesn't answer is said at the address field in the App language, the stored password is shown only as dots, the reolink-cli settings are folded under Advanced, and each Connect button names its service for screen readers (#1345).
  • An automation that says its answer at the door also tells you who came: its notification describes the visitor, then gives what was said at the door, instead of only the greeting (#1339).
  • A model's license dialog names the model instead of showing {{model}} (#1338).
  • An automation that a doorbell's button starts greets whoever rang, even when the picture shows no one, instead of saying nothing. A run that finds nobody to talk to notifies "Nobody was in the picture, so nothing was said at the door." instead of "Nothing". A visit's end mark is never said aloud, in any case (#1339).
  • Speaking and listening through a Reolink doorbell no longer stop when reolink-cli's gateway goes stale. For a gateway on this computer, Toskar starts it when it isn't running, and restarts it and tries again when it can no longer reach the doorbell.
  • The Access from anywhere plans in Settings show what the checkout charges, in its currency, as billing reports it, instead of prices written into Toskar. Until billing answers, the usual US prices show (#1257).
  • Live voice mode no longer jumps in while you're still talking. It waits for 0.9 s of quiet, not 0.5 s, before it answers, so a pause mid-sentence doesn't end your turn, and a cough, a knock, or a click no longer starts a turn or stops an answer.
  • An answer in voice mode is read in one voice from start to finish. A short sentence such as "Claro." no longer switches the voice or accent mid-answer in Spanish and other languages.
  • Voice mode's captions stay on screen until the next words replace them, instead of clearing when a noise is heard or the next answer starts.
  • Live voice mode no longer stops listening before you've finished a long sentence, or cuts off your last words, with a microphone across the room such as a webcam's. A turn now lasts through words that are softer than the ones that started it, and the room's noise is learned from your pauses rather than from your voice. The half second after your last word is now also sent to transcription, so a word that trails off is kept.
  • Live voice mode keeps Toskar's answers in the voice chat, so the next turn hears them and the chat shows both sides (#795). The answer was spoken but stopped before it was saved.
  • Live voice answers start well over a second sooner on a Mac, about 0.9 seconds from the end of a question to the first word heard, from 2.3 (#795). Each chat turn rebuilt what Toskar knows about the computer's models and hardware, looking at the hardware five times; it now keeps that between turns, and refreshes its look at the hardware in the background instead of making a turn wait for it.
  • Spoken answers are no longer asked for Markdown, which contradicted the spoken style (#795).

Security

  • Paired devices get a device token from the relay, which can look the computer up but can't open its tunnel or change its record, instead of the computer's own token.
  • A relay token handed over by a device is checked with the relay first, and can't replace a longer one.
  • Companion keys can't make companions of their own, and revoking a key revokes its companions.
  • The remote listener's lockout no longer turns away valid keys.
  • New API certificates don't carry the computer's name in their subject.

1.14.0

November 15, 2026

Downloads and release notes ↗Docs for 1.14.0 →

Added

  • A chat model can be kept running on a computer (#966, contract 1.94): PUT /api/v1/models/{id}/always-running. It loads when Toskar starts and again within a minute if it stops, such as after a crash, and isn't unloaded when idle or to make room for a picture or clip. A training run still frees the memory, and the model comes back after it. Stopping it yourself keeps it stopped until you start it again or Toskar starts. Turn it on in Models, under a model's More actions → Always running, or with toskarctl models pin. The model's card says Always running and how much memory it leaves for other models (memory_left_bytes, contract 1.95), and a model that then doesn't fit says which model kept running holds the memory. The Running tab and Performance mark a model kept running too (always_running on GET /models/running, contract 1.96), and only Admins see the menu item, as only they can change it. With paired computers, placement sends a model's requests to the computer keeping it running, and passes over a computer whose models kept running leave too little memory for it when another has the model. When nothing else can run a model that doesn't fit beside one kept running, an Admin can stop that one for the answer (make_room on POST /chat, contract 1.98); it comes back once the larger model is unloaded.
  • OpenAI's gpt-oss 20B and 120B in the model catalog. They think before they answer and use tools. Being a mixture of experts, each word uses only part of the model, so how well they fit and how fast they run is estimated from that part.
  • Qwen 3.6 35B-A3B in the model catalog: a mixture of experts that uses 3B of its 35B parameters for each word, and sees pictures.
  • Effort sets how long gpt-oss thinks, as the model's own low, medium, and high reasoning effort: briefly on Fast, at length on Thorough. From the API, reasoning_effort does the same.
  • The user guide's Pictures, clips, speech, and vision page is a full how-to: making and changing pictures, making clips and their settings, setting them up, asking about a picture or a video, and speech in and out, with screenshots of a picture being made, the Images and video tab, and a question about a screenshot, and an example CI runs (#893).
  • The user guide's Profiles page is a full how-to: seeing and making profiles, each of the editor's five tabs, Topics for keeping an assistant on one subject, strategies, and where chats run. Compare independent drafts and Get more careful answers say when each is worth the time, and screenshots show the Profiles page, the editor, and an answer whose drafts were compared (#897).
  • Check links in answers, in Settings (off by default): after an answer is written, Toskar asks each site it links to, up to six, whether the page opens, and names the links that didn't under the answer. A page that's gone, or a site that doesn't exist, counts as dead; a sign-in, a limit, or a slow site doesn't, and nothing counts as dead when no site answered, as when this computer is offline. Each check is listed in What left this computer as Link check. It never checks addresses on this computer or its network, answers from a profile that can't use the internet, or chats using anything marked this computer only. The saved answer has meta.link_check, and chat.links_checked says when it's added (#209, contract 1.97). In the answer itself, a link that didn't open is crossed out with a warning mark, and one that couldn't be checked gets a question mark, each with a tooltip. The Toskar apps and chat portals mark the same links; a portal reads the chat again once the check is done.
  • Quality results can measure a model at longer windows, such as 32K, 64K, and 128K tokens, with a prompt filling about three quarters of each: reading and writing speed, the wait for the first token, and the memory it took (contract 1.91).
  • Models › Running shows the memory each model really holds, measured from its process, against the computer's memory, where it showed the catalog's estimate. GET /api/v1/models/running gives it as resident_bytes, and quality results record it (contract 1.90).
  • Agents at once, per model: with Advanced mode on, a model's More actions menu on Models sets how many of a swarm's agents run on it at the same time, 1 to 4 (1 by default). llama-server starts with --parallel set to it and a context of the model's window times that number, so each agent keeps the whole window; the model uses more memory, and its fit and Auto's window count it. A number that leaves an agent less than 8K of context (MODEL_SLOTS_CONTEXT), or that doesn't fit the computer (MODEL_SLOTS_MEMORY), is refused. A loaded model reloads with it on its next request. Swarms run that many agents on the model at once, and the rest wait; a paired computer's models count as it's set there. PUT /api/v1/models/{id}/slots sets it, and each chat model installed on the computer says slots (contract 1.88, #1135).
  • TOSKAR_MTP=on lets a model with its own multi-token prediction head, such as Qwen 3.5 and 3.8's -mtp files, guess ahead with it, with no draft model. It's off by default: on Qwen 3.5 4B it was 38% faster on average, but changed 2 of 9 answers and cut one short.
  • Projects keep related chats together with what they share: instructions, files, a folder on the Toskar computer, a default profile and model, and a memory of their own. A project's chats follow its instructions, search its files, and remember what they learn for the project's other chats, not your others. In the web and desktop apps, Chat's history lists your projects above your chats, each opening to its chats, and New project makes one. A project's page has its chats with New chat, its instructions, files and linked knowledge, a folder (Admins), its default profile and model, and its memory switch. A chat says which project it's in, under its title, with Move to project and Remove from project. The API has them too (/api/v1/projects, and project_id on chats), so the apps can list projects and start chats in them (contract 1.87, #1136).
  • The phone and tablet apps list your projects above Computer in the chat list, with New chat in this project, and the TV and the car continue a project's chats; projects are still made and changed on the computer (toskar-apps#432, #1136).
  • The Memory page lists what a project's chats learned under the project's name, after your own memories, and Move to personal memory in a memory's menu makes it yours in every chat. The API does the same with PATCH /api/v1/memory/{id} and an empty project_id, and a memory says which project it belongs to with project_id (contract 1.87, #1136).
  • In the terminal, toskarctl projects lists your projects with how many chats are in each, and toskarctl chat --project <name or id> starts a chat in one, or with --resume continues one of its chats. The chat still works in the folder you're in, before the project's. While chatting, /project says which project the chat is in, /project <name> moves it into one, and /project none takes it out (#1136).
  • The user guide's Projects page is a full how-to: making a project, its page, moving chats in and out, what it remembers, and projects in the terminal and the API, with screenshots of History, a project's page, a chat in a project, and the Memory page, and examples CI runs (#1136).
  • Details on Models › Results shows where a result ran (its graphics card, llama.cpp release, guessing ahead, and memory) and, when the run measured them, the same model at longer context windows: how fast it read and wrote, how long it took to start answering, and its memory.
  • seed, top_p, top_k, and min_p on /v1/chat/completions and POST /chat, which also takes temperature now. They reach the calls that write the answer, here or on a paired computer. The same seed and settings usually repeat an answer from one model; see docs/api.md for what is and isn't promised (#839, contract 1.93).
  • Swarms for coding agents: swarm:<profile> as the model on /v1/chat/completions, /v1/messages, and /v1/responses. A request that brings the agent's own tools, as a subagent's does, goes as it is to the least busy of the profile's swarm models and computers, on this computer or a paired one, and its answer comes back unchanged, so subagents started at once run at once instead of waiting on one model. A request without tools runs a whole swarm and answers with its answer, with the agents' progress as reasoning. /v1/models lists swarm:<id> for profiles whose strategy is swarm (contract 1.88, #1135).
  • toskarctl connect claude-code --swarm <profile> points Claude Code's subagents at a profile's swarm, and leaves its own model as it was.
  • MCP's run_swarm tool runs a swarm from another app and returns its answer and a link to its chat.
  • Swarms for tasks in a folder: when a task in a chat's folder asks for a swarm, the coordinator looks at the folder, read-only, and splits the task into parts with the files each expects to change; parts that change the same files go to one agent. In a Git repository each agent works as a task in a worktree of its own, on a branch of its own, starting from what you see, uncommitted changes and new files included, and its work is committed there as Toskar. The branches are merged one at a time; a conflict goes to a fix-up agent, and one it can't settle is reported with the agent's branch. The project's checks run on the merge, and the result is applied to your folder as uncommitted changes once you allow it. A folder that changed since the swarm started, so the change no longer applies, isn't forced: the answer names the branch to merge. In a race, an agent whose checks fail can't win over one whose checks pass, and the judge picks among the rest. Your branch, HEAD, index, and stash are never touched. Worktrees are removed when the swarm ends; branches are kept for 7 days, or until DELETE /api/v1/swarms/{id} deletes them. In a folder that isn't in Git, the agents take turns in the folder itself, and the answer says so. Swarms gain folder, branch, checks, conflicts, and applied, and their agents paths, branch, and checks (contract 1.88, #1135).
  • Swarms in the terminal: toskarctl chat --swarm sends each message to a swarm (split, the default, or race), and /swarm sends the next one, with /swarm off to stop. While the agents work, each has one line, updated in place: its part, computer, model, how it's doing, and its latest step; then the merge, any change that couldn't be merged, and for a task the branch, the checks, and whether the change is in your folder, then the answer. Piped, each change is a plain line on standard error. An agent's tool questions show and are answered in the terminal, labeled with the agent, and "always" applies to that agent alone. /agents lists the last swarm's agents, /open N continues an agent's chat, and /back returns. Ctrl-C stops the whole swarm (#1135).
  • Swarms on the web and desktop: Chat options has Use a swarm, Split or Race, for the next message, shown as a chip you can remove. While a swarm works, the chat shows a lane for each agent with its part, computer, model, status, latest steps, and result, and Stop and Open for that agent; then the merge, any conflicts, and in a folder the result's branch, checks, and whether it was applied. After the answer, the lanes stay above it, folded. Agents' tool questions show in the chat that started the swarm, each named by its agent, and Allow for this agent covers that agent alone. Profiles gain the Swarm strategy, with its mode, models in order, and agent, token, and minute budgets. A project's page folds a swarm's agents' chats under the chat that started it, and an agent's chat links back to it (#1135).
  • Swarms: several agents answer one question at the same time, each in a chat of its own. In a split, the question is split into parts and the answer is written from every agent's; in a race, agents on different models each answer the whole question, and the best answer wins or they are combined. Agents run at once on different computers or different models; one computer with one model runs one agent and says why. Ask for one with "use a swarm" or "use 3 agents" in a message, swarm on POST /api/v1/chat, POST /api/v1/swarms, or a profile's new swarm strategy, which Auto never picks. Agents' chats go in the chat's project, or a new one named after the question; GET /api/v1/conversations leaves them out unless swarm_id is given. Budgets for agents, tokens, and minutes stop agents still running and merge what finished. An agent's tool questions say which swarm and agent ask, and allowing a tool for the session applies to that agent's chat alone. Stopping an agent's chat stops that agent; stopping the swarm stops them all (contract 1.88, #1135).
  • The user guide's Swarms page is a full how-to: split and race and when each helps, starting one from Chat, a profile, a message, the terminal, VS Code, coding agents, MCP, and the API, the lanes, tasks in a folder with their branches, checks, and applying, agents' approvals, budgets, Agents at once, which keys and roles each way needs, and what the phone, tablet, TV, car, and watch show. It has screenshots of Use a swarm, a profile's swarm, the lanes, a task's result, an agent's approval, Agents at once, and a project's agents, and examples CI runs (#1135).

Changed

  • A mixture-of-experts model too big for the graphics card is labelled by the computer's memory, with its experts running from it and the rest on the card, instead of Tight like a dense model of its size.
  • gpt-oss 20B can check a profile's topics for other models: it held every topic case in the quality set, against 20 of 22 for the smallest checker before it. Toskar still uses the smallest checker that fits.
  • The quality run measures swarms against a single agent (#1135): six questions in several parts, answered from the set's fixed web pages, and four tasks in a Git folder whose parts are in files of their own, scored by the folder's tests. Each is asked of one agent on each model and of a swarm on the same models (split, and race with two or more), and the report counts what was done, the time, the tokens, how many agents ran at once and on how many computers, and why a swarm ran as one agent, with the swarm against one agent side by side. The set answers yes to writes in its own folders and to applying its swarms' changes there, and no to everything else. The weekly run measures the recommended models with two slots, and the Quality workflow's sets, swarm_models, and swarm_slots inputs run it by hand.
  • The Runtimes page names llama.cpp's release, such as b12361, for a llama.cpp Toskar installs from now on, where it showed the date of the file.
  • Quality results say which llama.cpp release a model ran on, whether a draft guessed ahead and how many of its guesses were kept, and every GPU it ran on (contract 1.89).

Fixed

  • Asking for code that works on a file, such as "write a Python script that reads a CSV of orders", gets the code. Toskar made the CSV file instead.
  • Toskar keeps its database in its data folder when the folder's path has a # or ? in it. It opened one beside the folder instead, and folders whose names began the same shared it.
  • gpt-oss answers when a tool it's offered takes a list, such as making a spreadsheet's rows. Its chat template stopped with "Function is not a bool value" on a list that didn't say what it holds.
  • A file asked for in chat holds what the model wrote, when the model made the file with its own tool call, as gpt-oss does. The file held the call's text instead, or nothing at all.
  • gpt-oss writes files again. Toskar sent its tools' settings in alphabetical order, and the model's file writes came back with the file's contents cut off, so tasks ended with the fix described but not made. They now keep their own order.
  • Questions about Toskar itself are answered from the user guide's detailed pages, not only their overviews. Asked what a profile's Never discusses field does, chat said there was no such option, since only the Profiles page's body describes it (#1222).
  • The user guide said Profiles & Orchestration needs advanced mode and its editor has four tabs. Anyone can open it; editing needs Advanced mode, and the editor has five tabs, Topics included (#897).
  • Knowledge no longer brings a passage into a chat because it shares one everyday word with a longer question. Asked about coffee shops near the harbor, a refund policy for "Harbor Supply" came along, and a small model said it had no information instead of looking it up. A shared code or size, such as a SKU, is still enough on its own, and passages that turn out to be about something else are set aside in the answer (#1217).
  • On Linux, a running model's measured memory is what its process allocated. It counted the model's file too, which stays mapped after the weights go to the graphics card, so the same model read anywhere from 2.8 GB to 9.3 GB.
  • A small model that numbers its tool calls, such as {"tool_call":{"id":"1",…}}, gets its call run when only one tool fits it, or is shown the call with each tool that fits so it picks one. A turn that still ends on a call is answered without tools. Before, Qwen 3.5 4B asked about coffee shops repeated the numbered call until the turn ended with "The model didn't write an answer" (#1213).
  • Offline, a question that needs the internet, such as today's AI news or a web search, is answered with saying this computer isn't connected. Before, the model wrote an answer from memory as if it had looked: asked for the news, it made up the day's stories (#965).
  • Pages and setup steps that need the internet say this computer is offline, and which site couldn't be reached, instead of showing dial tcp: lookup huggingface.co: no such host or sending you to Diagnostics. API errors for it have the code OFFLINE (contract 1.92).
  • Asking for a kind of picture, such as "make a watercolor of a lighthouse at dusk", a sketch, or a cartoon, makes a picture, or offers to set up picture making. Before, it was taken as a question about painting and answered from the web (#893).
  • The user guide's Pictures page says what the Tools page shows about each model, Recommended for this computer or how much memory it needs, and the real download sizes (#893).
  • Asking for a logo, icon, portrait, poster, wallpaper, or pic before picture making is set up offers to set it up, as asking for an image does. The three places that read a request for a picture now share one list of kinds of picture, so they can't drift apart again (#1225).
  • Profiles & Orchestration says why an action was refused, such as a name another profile has. Before, saving, renaming, creating, duplicating, resetting, deleting, and choosing the default failed without a word, and a refused name just stayed as it was (#897).
  • A computer stays reachable away from home when its subscription renews, without opening the app on the phone that bought it. Before, it dropped off when its old access ran out, or after a restart.
  • A model bigger than the computer's graphics card shows no speed on the Models page until it has run there. Its estimate was up to ten times too fast: such a model runs partly on the processor.
  • A swarm's agents find their own part in the profile's knowledge, and each works on its part alone. An agent got another part's files, and said the knowledge didn't have its facts.
  • A task, or a swarm agent's part, that names a file, such as "Write Reverse in strs/strs.go", changes that file in its folder instead of attaching a new file to the chat. Ordinary chats with no folder still make the file (#1135).
  • A task swarm's change whose project checks failed, or ran out of time, always asks before it's applied to your folder, whatever your profile allows. The question says the checks failed, and the answer says whether the change was applied because you chose to or wasn't, with its branch (#1135).
  • In the terminal or a chat with a folder, a request to fix or change code later in the message, such as "The test in sum_test.go fails. Fix Sum so it passes.", is done as a task instead of answered as a question. "Why does my test fail?" is still a question (#1135).
  • POST /api/v1/chat/stop stops only the caller's own chats; another person's chat reads as not running.
  • Writing files, running terminal commands, and Git add and commit ask first again in the built-in profiles, and Allow for session still lets them run for the rest of a chat. Git push asks every time, since it sends your code off the computer. Before, General Assistant and Programming let all of them run without asking (#1206). In a task, commands that only read or test, such as go test, still run without asking, so a fix-and-test loop can.
  • Settings → Tool permissions now takes effect. Each switch is a limit for every profile on this computer: a profile can be stricter, never looser, and Deny holds in automations too. Before, nothing read them (#1206).
  • The AI is told that tools ask first only when they do, and is told to change only what was asked when some run without asking (#1206).

1.13.1

November 11, 2026

Downloads and release notes ↗Docs for 1.13.1 →

Fixed

  • The guide's Use Toskar in VS Code covers disconnecting, a key from API Access where nobody can type a code, models that Agent mode leaves out, and the certificate error earlier versions of the extension showed after the code.
  • Models installed from Hugging Face or by address now say whether they call tools and think, with their family and context window, read from the model file. Models already installed this way are read once at start, so a model such as Qwen 3.8 shows up in editor model pickers that list only models that call tools.

1.13.0

November 7, 2026

Downloads and release notes ↗Docs for 1.13.0 →

Added

  • Text for the app's Access from anywhere card when another device's subscription already covers the computer, and when a computer has more than one subscription (yeixio/toskar-apps#388).
  • Toskar Pro from the Mac App Store sells access from anywhere in API Access, as an App Store subscription at the App Store's prices, with Restore purchase and Manage subscription (yeixio/toskar-apps#245). Toskar Pro from the Microsoft Store, like every other copy, keeps the checkout billed by Yeix.io LLC.
  • A chat can have a folder on the Toskar computer to work in. In one, work asked for goes on until it's done on the web and desktop too, as it does in the terminal, and a task started in the terminal keeps its folder when it goes on elsewhere. Choosing one needs the admin role (contract 1.85, #1017).
  • Ask for the AI news or your automations' latest results in any chat, such as "read me today's AI news" or "what's in my briefing" in the car, in any of the app's ten languages. Each person hears only their own automations, and with AI news off, chat says it's off in Settings (toskar-apps#142).
  • toskarctl chat starts Toskar when it isn't running on this computer, through its service if it has one, and says how to stop it. Toskar stays running after the chat. If Toskar isn't installed, chat asks first, shows the download's size, and runs the installer. With TOSKAR_URL set, it starts nothing.
  • A step-by-step guide for Claude Code and Codex CLI with Toskar as their model provider, on this computer and from another one with a key, tested with Qwen 3.5 9B (docs/integrations/coding-agents.md, #1060), linked from the user guide's Terminal page.
  • A step-by-step guide for Continue in VS Code with Toskar as its model provider, on this computer and on another one with a key, tested with Continue 2.0.0 (docs/integrations/continue.md, #1061), linked from the user guide's Terminal page.
  • curl examples for following Toskar's events (examples/curl/events.sh) and reading this computer's hardware (examples/curl/hardware.sh), checked with the other examples (#816).
  • Profiles marks the profile new chats start with, and More actions → Use for new chats makes another one the default. Settings' Default profile shows the same, with no Automatic choice.
  • The user guide's API page is a full how-to: finding the endpoint, API keys and what each may use, the Python, TypeScript, and Go SDKs and the client contract, toskarctl ask in scripts, and the GitHub Action for CI, with a screenshot of API Access and examples CI runs (#898).
  • The user guide's Knowledge page is a full how-to: connecting a file or folder, seeing sources in an answer, connecting a database or a web API, and keeping knowledge fresh, with screenshots of the Knowledge page and the database form, and examples CI runs (#890).
  • A user guide page, Use Toskar in the terminal: toskarctl chat, handing it files and command output, working in a folder with approvals, Toskar on another computer, and toskarctl connect for coding agents, with examples CI runs (#959).
  • The user guide's Tools page is a full how-to: seeing your tools, choosing when each asks first, answering a tool that asks, adding tools from an MCP server, and using Toskar from other AI apps, with screenshots of the Tools page and an approval in chat, and an example CI runs (#892).
  • The Toskar extension is on the Visual Studio Marketplace and Open VSX, and the user guide has a page for it: Use Toskar in VS Code.
  • A task in the terminal now keeps its own record of what it checked and changed. If the last check failed, nothing ran to check a change, or test files changed, the model is sent back once, and whatever is still in doubt shows under the answer (#1017).
  • Web and desktop chats can do long tasks. An Admin chooses a folder for a chat from the composer's options, and the chat then shows each command's output as it's written. A task started in any app shows in the others: its steps, its output, and its questions, which any of them can answer, and any can stop it (#1017).
  • GET /conversations/{id}/turn says what a running turn has done so far, and tool.decided says a waiting tool was answered (contract 1.86, #1017).
  • A chat with a folder names the model measured to finish tasks, such as Qwen 3.8 27B, under Working in until one is installed, with its size, how well it fits, and a link to the Models page, as toskarctl chat already did. On the quality runs, Qwen 2.5 7B and Gemma 2 9B finished none of the long tasks, and Qwen 2.5 32B most (#1017).
  • Text for a TV that can buy Toskar Unlock itself, saying it can be unlocked there or on a phone or tablet (yeixio/toskar-apps#134).

Changed

  • In the car, someone without Toskar Unlock hears what it adds, and that the phone will offer it once they're parked (yeixio/toskar-apps#398).
  • toskarctl chat with no model installed now offers the one that suits this computer and installs it, instead of an error. While a task works, each step is a line in plain words, with a line when a command fails or takes a while (contract 1.84, #1017).
  • Toskar recommends Qwen 3.5 9B on most computers with 16 GB or more, and Qwen 3.5 4B with an 8 GB graphics card, instead of Qwen 2.5 7B. On the long tasks, Qwen 3.5 9B finished all three, where Qwen 2.5 7B finished at most one. Onboarding also skips a model that only just fits, as Best overall does, so 8 GB computers keep Llama 3.2 3B.
  • A task in the terminal now asks before it changes anything, whatever the profile allows for chat: writing files, git add and git commit, installs, make and npm run, and reading outside the project. Reading in the project and running named tests and builds still don't ask (#1017).
  • In a terminal task, a command can run for up to 30 minutes, with its output showing as it comes. Ctrl-C stops the work and stays in the chat. Stopping a command now stops what it started too, and something left running in the background no longer holds up the answer (contract 1.83, #1017).
  • The user guide's Terminal page shows a recorded session: Qwen 3.5 9B finding a failing test's cause and asking before it changes the file. CI runs a whole toskarctl chat session against a real Toskar with the scripted model (#959).

Fixed

  • toskarctl looks for Toskar at TOSKAR_API_PORT when that's set, the same port Toskar itself uses.
  • Claude Code and Codex work with Qwen 3.5. Toskar passed their system messages on where they were sent, partway through the conversation, and Qwen 3.5's chat template refused them, so every request failed with "System message must be at the beginning". They now reach the model as one system message at the start (#1060).
  • toskarctl connect continue adds useLegacyCompletionsEndpoint: false to the model it prints, so Continue's edit tool and Apply go through chat instead of /v1/completions, which Toskar doesn't have. Before this, they failed. With TOSKAR_API_KEY set, the key is read from Continue's ~/.continue/.env (#1061).
  • New chats start with the default profile, General Assistant unless you choose another. A profile used once with Use in Chat, or opened in another chat, no longer carries over to every new chat, and a paired computer no longer switches new chats to a Team profile. Chats started without a profile, from the phone, another app, or the API, use the default too.
  • Profiles can't share a name. Setup now fills in the built-in profile for its purpose rather than making a second General Assistant, and profiles that shared a name are renamed once, such as General Assistant (2), with none deleted.
  • In a chat with a folder, or a task in the terminal, Git works in that folder, as the file tools and commands do. It used Toskar's own workspace, so a commit could land in the wrong place. Git's view of the project no longer counts as outside content either, so a task that checks git status first doesn't ask before every later step (#1110).
  • gpt-oss can use tools and do tasks. Its calls come back from llama.cpp only when the request offers the tools itself, so a chat with a folder, a web search, or any other tool stopped before doing anything. Toskar now offers them that way to models in gpt-oss's chat format, and reads the call as any other (#1099).
  • The user guide says that traffic between paired computers, and from devices on your network, is encrypted. It still said "not encrypted yet".
  • A model added from a file can use tools when its chat template takes them, as gpt-oss, Qwen, and Llama do. It was treated as unable to call any, so its answers to ordinary questions, such as one about a file in the chat's folder, could only guess. Models added before this are read from their file, with nothing to add again (#1101).
  • A question that uses all its tool calls is answered from what they found, instead of "I stopped after 3 tool calls so this request would not loop".
  • On Linux without systemd, such as a container or WSL without it, the install script no longer waits a minute and reports a failed install that worked: it says nothing keeps Toskar running there and how to start it. toskarctl chat starts Toskar itself when the installer it ran put Toskar in place but couldn't start it.
  • A request for places of a kind, such as "Find a cafe in downtown Juneau", is looked up on the web before the model answers. Models that don't search by themselves, such as Qwen 3.5, made up names and addresses (#1126).
  • Qwen 3.5 models use tools. They put their own label for a call in its id, such as terminal_1 or run_test_001, so every call was refused and they answered without tools. Toskar now reads such a call as the tool it names, or the one tool its arguments fit, and asks again by the tools' ids when it can't tell. On the long tasks, Qwen 3.5 4B went from none done to 2 of 3 (#1124).
  • Benchmark results sent from the app go into the public results by themselves again, when their model's name was sent as its id. The results page now shows each model under its name from the model list, never one a sender wrote, and the benchmark sends the list's name for a model installed from Hugging Face (#1087).
  • Home CI runners download each model once and keep it, and stop sending caches to and from GitHub. Jobs report their network use in the summary.
  • In a chat with a folder, or the terminal, a follow-up that goes on with the work, such as "go ahead and make that fix", "yes, do it", or "try again", works as a task, with the task's tools. It was answered as a question, so the model could only describe the change again (#1111).
  • In the terminal and in a chat with a folder, a change now asks however the message is worded. A message that didn't read as a task, such as "Use the terminal tool to...", ran known commands such as touch without asking when the profile allowed them (#1017).
  • git stash, git switch -f and --discard-changes, git commit --amend, git pull --rebase and --force, and git fetch --prune no longer run without asking in a chat that allows known commands. A new set of attempts to get around the terminal's safeguards runs in CI and found them (#1017).
  • Qwen 3.5 can change files in a task. Writing a whole file, it closed its call one brace short, so Toskar dropped the request without asking, and the model said it had made the change. A call left open at its end is now closed, unless it was cut off partway (#959).
  • Finding files understands patterns such as *test* or *.go, as models write them. Before, it looked for those characters in the name and found nothing (#959).

1.12.3

November 3, 2026

Downloads and release notes ↗Docs for 1.12.3 →

Added

  • Text for the phone and tablet app's speed test, in every language.

Changed

  • The one-line benchmark can measure every model that fits, or only a computer's Toskar Score, in minutes. It keeps the computer awake, picks up where it stopped, and removes each model it downloaded.
  • The one-line benchmark offers the most popular model that fits, one nobody has measured yet, anything else on the list, or any model by its Hugging Face link.
  • The one-line benchmark can send its result with no GitHub account. Results also say which operating system they ran on.
  • Results people send are checked automatically: the ones that pass can go in the public results, and anything unusual waits for a person.
  • The Toskar Score is measured on Qwen 2.5 3B, which phones and tablets can run too.

Fixed

  • The one-line benchmark asks before downloading on Windows. It could take no answer for a yes, and start by itself.

1.12.2

November 3, 2026

Downloads and release notes ↗Docs for 1.12.2 →

Added

  • toskarctl chat has /log: what ran in this chat, its result, and who allowed it. The record keeps a command whole, and a command that fails counts as failed.
  • curl -fsSL https://toskar.ai/benchmark.sh | sh, or irm https://toskar.ai/benchmark.ps1 | iex on Windows, measures a model on your own computer, with nothing else to install. It suggests a model that fits, shows a Score and how fast it ran, and you choose whether to share it.
  • toskarctl help lists every command by what it's for, and toskarctl help <command> or --help explains each one, with its flags and examples that are tested. A mistyped command suggests the nearest one.
  • GET /api/v1/models/store: the model store's shelves for this computer, from Hugging Face's rankings and the models tested with Toskar.
  • docs/quality.md explains how quality is measured, and how to suggest changes. Results now name the version of the method that measured them.
  • scripts/quality-score.sh measures a model on your own computer, in a throwaway Toskar, and writes a result you can send us.
  • The benchmark has a Long prompt workload: a document of about 4,000 tokens, new each run, to show how long a model takes to read it and start answering.
  • Models has a Results tab: what Toskar's test runs have measured so far, per model, computer, and date, to sort through. It links to how it's measured, and how to add your own.
  • toskarctl chat works in the folder you start it in, on the computer Toskar runs on: commands run there, its file tools reach only that folder, and it follows the project's notes in AGENTS.md, TOSKAR.md, or CLAUDE.md. Reading a file in the project no longer makes the next known command ask after "always".
  • In toskarctl chat, work you ask for now goes on until it's done or stuck, instead of stopping after a few tool calls, with the tools it may turn out to need. It says what it did and what's left when it stops.

Changed

  • Terminal commands that can delete or change things outside the chat, or that Toskar can't read or doesn't recognize, now ask every time. This holds even for a profile that allows the terminal and after "always in this chat", and automations skip them and report it. A command that can't be undone says so.
  • toskarctl chat shows answers as they're written, and rewrites what it showed when Toskar changes it, such as taking back a sentence before using a tool. Piped output still gets the whole answer.
  • After "always in this chat", output from a known command such as go test no longer makes the next known command ask, so a fix-and-test loop can keep going. Web pages, files, and commands Toskar can't vouch for still make it ask, and the prompt says what "always" trusts.
  • The Models page is now a store: shelves of popular, top rated, and new models from Hugging Face, plus everyday, code, pictures, small, and tested by Toskar, all of which run on your computer. Choose a model for its license and a link to its Hugging Face page.
  • The public model results only hold complete rows, with all of Quality, Topics, Files, and Reasoning, so every row can be compared.
  • Model results name a computer the same way in every run: its graphics card or chip, and the memory a model can use there. They also record what a model ran on, such as Vulkan or Metal.
  • Releases include toskar-quality, which measures a model on your own computer without Go, Python, or a checkout. Its result now has a Score: the average of the four sets' pass rates, shown with the method's version.
  • The weekly quality run works through a public list of popular models, a few each week, so the results slowly cover what people use.
  • Model results record how a model ran: reading and writing speed, first token, start-to-finish speed, GPU layers, GPU memory, quantization, and context window.
  • Model results say whether a model runs fully, partly, on the processor alone, or not at all on that computer. A model that doesn't run gets a row of its own, with no scores.
  • Each weekly quality run adds its scores to the public results by itself.
  • The Results tab on Models shows each result's Score, and how fast the model read and wrote.
  • Recommended on the Models page now shows a model we've tested, then the most popular models that run well on your computer.
  • Qwen 3.8 27B now counts as good for code, so recommendations and toskarctl connect can pick it. In tests it fixed a failing test where the smaller coding models mostly couldn't.
  • toskarctl chat says which model to install when nothing installed does well at work such as fixing code.

Fixed

  • After a chat reads something from outside, such as a web page, a tool that changes things asks again even if you answered "always in this chat". Before, only a profile that allowed the tool outright was held to this.
  • toskarctl models install no longer gives up after a minute while a large model is still downloading.
  • toskarctl chat --model and /model say when a model isn't installed, rather than let another model answer without saying so, and a chat shows the answer's notice when another model answered.
  • A benchmark of a model that thinks first reports its real time to the first token, not zero.
  • Asking about a video, such as whether its frames change, no longer makes a new clip instead of answering, and someone who can't make videos can ask about one. A clip's file name no longer picks up the model's tool call text.
  • A quality score counts every Topics case as its report does, so a wrong refusal lowers it.
  • The terminal tool works on Windows without sh: commands run in PowerShell, and the model is told so.

1.12.1

November 1, 2026

Downloads and release notes ↗Docs for 1.12.1 →

Added

  • The chat API can send an answer as it's written. A client that sends "live": true with "stream": true gets the text as the model writes it, and a replace event when a tool call, a retry, or a check changes what was shown. Without it, the answer still comes whole (contract 1.77).

Changed

  • Chat on the web and in Toskar Pro shows an answer as it's written, instead of all at once when it's ready. The text can change or start over before it finishes, when Toskar looks something up or corrects a figure.

Fixed

  • The curl chat example sends messages with quotes, backslashes, tabs, or several lines; it used to send invalid JSON for them.
  • Deliberate's chosen answer is what you get. With answer checks on, which is the default, the turn's first draft was shown instead.
  • Wan 2.1 T2V 1.3B makes clean 2-second clips on Apple silicon. At its old settings they came out with harsh color or broken frames. A clip takes about half as long again as it did.

1.12.0

November 1, 2026

Downloads and release notes ↗Docs for 1.12.0 →

Added

  • The phone app in CarPlay asks as client: car (contract 1.65). Like the watch, it gets a short answer from whatever model would answer best, and a model it used stays loaded for an hour, so the next question starts at once. The phone app's CarPlay text, and its line in Toskar Unlock, are in all ten languages (yeixio/toskar-apps#142).
  • A chat can turn a reasoning model's thinking off, for answers that start straight away. With a model that thinks before it answers, the chat's options have Thinking, On or Off, and the choice is remembered. `POST /chat takes "thinking": "off", as /v1/chat/completions` takes reasoning_effort: "none" (contract 1.61, #840).
  • toskarctl connect also covers editors: continue, cline, roo-code, zed, jetbrains, and copilot print what to enter in that editor's settings for its agent or AI features to use a model on Toskar, with Zed and Cline told the model's context window (#960).
  • A request that needs a bigger context window gets one. On /v1/chat/completions, a request that names a model and needs more than the model runs with makes it start again with the smallest size that holds it, up to its own window, if it still fits. Coding tools can ask up front with toskar.context_window. When nothing fits, the request answers 400 CONTEXT_TOO_LONG naming the field. The model goes back to its usual window once it stops (contract 1.70, #962).
  • Choose a model's context window. With Advanced mode on, a model's More actions menu has Context window: Auto, or a size from 4K up to what the model was made for, each marked with how well the model fits at it. A bigger window suits long documents and coding tools; a smaller one frees memory. A running model reloads with it on its next message. PUT /models/{id}/context sets it, and models list context_window, context_choice, and context_options (contract 1.69, #962).
  • Run traces mark chats from @toskar in VS Code with the new editor client (contract 1.75) (#960).
  • TOSKAR_KV_CACHE=q8_0 starts chat models with an 8-bit key/value cache, about half the memory for the same context window, so the quality set can measure it before it's a default (#962).
  • A nightly llama.cpp canary checks that the newest release, the one Toskar installs, computes what the known-good tag does: Metal and the CPU on a Mac, Vulkan and the CPU on Linux. It compares next-token probabilities over a fixed text (KL divergence, top-token agreement, perplexity) and opens an issue with the numbers when a release drifts (#833).
  • On the Mac, the canary also checks a model of each family that thinks before it answers (Qwen 3, DeepSeek R1's distills, gpt-oss): the thinking comes apart from the answer, and turns off when asked (#840).
  • Coding agents that speak Anthropic's Messages format can use a model on this computer. POST /v1/messages converts their system prompt, tool calls, tool results, pictures, and tools to the model's format and the answer back, streamed as Anthropic's events or not, and POST /v1/messages/count_tokens estimates a request. It's raw, with the same permission as raw:<model>, and takes the key as x-api-key too (contract 1.72, #843).
  • Models: Qwen 3.5 4B, Qwen 3.5 9B, and Qwen 3.8 27B, which also see pictures. They answer first and think only when asked. Qwen 3.5 4B replaces Qwen 2.5 3B in the General Assistant's choices: on the rig it got 39 of 39 reasoning questions right, against 28 (#593).
  • An editor can connect like a device, with the code from Connect a device, and ask for raw access for its coding agents ("raw": true on POST /devices/pair, contract 1.76). It gets it when an Admin or the Owner showed the code (#960).
  • Diagnostics shows how much of a chat's later turns each model skipped reading again, from its prompt cache, over the last week. A model that reads most of a long chat again every turn is marked low, since its answers start later as a chat grows. GET /diagnostics/prompt-cache gives the same per model (contract 1.62). The user guide's new page, Why do answers in a long chat start slowly?, explains it (#838).
  • Coding agents and editors can drive a model directly. On /v1/chat/completions, "model": "raw:<model id>" sends the request to a model on this computer as it came, with the agent's own tools, tool_choice, stop, and logprobs, and returns the model's tool_calls, streamed or not, without Toskar's assistant. Its window is sized as other requests' are. A request from this computer may do this without a key; for a key, an Admin turns on Drive models directly, for coding agents, in its permissions (contract 1.71, #843).
  • Coding agents that speak OpenAI's Responses format can use a model on this computer. POST /v1/responses converts their instructions, input items, function calls and their outputs, pictures, and function tools to the model's format, and the answer back as output items, streamed as the response.* events or not. It's raw, with the same permission as raw:<model>. Toskar keeps no responses, so an agent sends the whole conversation each time (contract 1.73, #843).
  • Draft models: a catalog model can name a small model of its family as its draft, downloaded and deleted with it and counted in its memory fit. Core starts the model with the draft and guessing from the chat together when the model fits the computer well (Good or Excellent), and Guess ahead isn't off for it. Qwen 2.5 32B and 14B come with Qwen 2.5 0.5B, which made them about 1.9× and 1.6× as fast on an RX 7900 XTX, changing answers no more than the GPU does between two plain runs (#835).
  • Guess ahead, Auto or Off for each model: with Advanced mode on, a model's More actions menu on Models turns speculative decoding off for that model, from its next start. PUT /api/v1/models/{id}/speculation sets it, and each chat model installed on the computer says speculation (#835, contract 1.67).
  • Performance shows how often speculative decoding's guesses pay off: with Advanced mode on, a reply's details say how many words the model guessed ahead and kept, and Activity shows the share kept across the replies shown, per model when filtered by one. GET /api/v1/performance has each reply's draft_tokens and draft_accepted (#835, contract 1.66).
  • A coding agent with many tools, such as one with MCP servers, works with a local model. When a raw request's tools take more than a quarter of the model's window, the model gets the most relevant ones and a tools_search tool to find the rest; Toskar answers it, adds what it found, and asks again, so the agent sees only calls to its own tools. A long tool list no longer makes a request too long for the model (#843).
  • toskarctl chat chats with Toskar in the terminal, in the same chats as the apps: answers stream, --resume continues a recent chat, a tool that asks first waits for y or n, and piped input is answered once as data. Run traces mark these chats with the new terminal client (contract 1.74) (#959).
  • toskarctl connect claude-code, codex, opencode, or aider prints the settings for that coding agent to use a model on Toskar, with its own tools, and --write saves them: for Claude Code and OpenCode in the project folder, and for Codex as a toskar profile, so the agent's usual settings stay. It picks the best installed model for coding unless --model names one (#959).
  • toskarctl models list lists the models, smallest first, with their download size, fit on the computer, and whether each is installed (--installed, --json). toskarctl models install <id> downloads one and waits until it's installed, printing its progress, so a computer without a desktop can add a model from a terminal (#887).

Changed

  • A model that thinks before it answers gets more room for its answer when a chat's earlier messages are fitted into its window: a quarter of the window (2,048 to 16,384 tokens) instead of a flat 512, or what the request's max_tokens asks for, and the usual room when thinking is off. Other models get a sixteenth of the window (512 to 4,096) (#840).
  • Toskar knows a model thinks before it answers from the model file: its chat template, or its name for DeepSeek R1's distills. A Qwen 3 or gpt-oss added from a file or from Hugging Face counts, and a model that's only good at reasoning, such as Qwen 2.5 32B, doesn't. Models carry thinks in their capabilities (contract 1.63, #840).
  • /v1/chat/completions checks that a request naming a model fits it before running: max_tokens as large as the model's window, or messages and max_tokens clearly larger, answer 400 with CONTEXT_TOO_LONG and the field to change, without starting the model (#840).
  • A catalog model can answer first: it thinks only when a chat asks it to. Smaller reasoning models such as Qwen 3.5 spent their whole window thinking and wrote no answer. For these models the chat's Thinking option starts Off, and the choice is remembered apart from the other models'. POST /chat takes "thinking": "on", and any reasoning_effort other than none asks them to think. Model capabilities have answers_first (contract 1.68, #593).
  • Catalog: each model names its family (its line, such as qwen) and its generation (such as 2.5), and can name the entries it supersedes, size for size. A test fails when a model has no family or generation, or supersedes one of another family, the same generation, a different job, or itself by way of others. Models still show their family as before, such as qwen 2.5 (#593).
  • Chat models start with a bigger context window where it fits. Toskar used 8,192 tokens for every model; now each starts with 8,192, 16,384, or 32,768, the largest at which it's still an excellent fit and stays in graphics memory, and never past the model's own window. Small models on most computers get 32,768, so they read about four times as much at once, and a reasoning model has room to think. A model that only just fits keeps 8,192. The fit estimate in Models counts the bigger window, and a model running with another window is started again (#962).
  • When the nightly llama.cpp canary finds a release computes differently from the known-good one, new installs take the newest release before it until a newer one passes (runtimes/llama-cpp.json). When the hold can't be read, installs take the newest release, as before (#833).
  • With paired computers, a chat goes to the computer expected to answer soonest. Toskar learns how fast each one reads prompts and writes answers with each model from the answers it gives and from benchmarks, and counts the work in flight on each, so a long request goes where prompts are read fastest and a short one where answers are written fastest. A chat's next message stays on the computer that has the chat loaded unless another would be much faster. The placement reason shows each part (#836).
  • The prompt cache check also asks, warm and cold, which city the store is in, which only the chat's first sentence says. A cache that lost the start of a chat answers another city, so this catches it on every model, Gemma 3's sliding window included (#838).
  • The weekly prompt cache check also checks that a new chat reads its system prompt from the cache, even after a chat with another one, as llama-server's in-memory prompt cache allows. A profile's long instructions are read once, not per chat: with Qwen 2.5 7B, 88 ms instead of 1.6 s for 2,100 tokens (#838).
  • The weekly prompt cache check also runs Gemma 3 4B, whose sliding window is the catalog's smallest, so a llama.cpp update that stops reusing such a model's cache fails it. Its chat is now about 2,300 tokens, longer than that window, with the model writing its own answers, and warm and cold answers are compared by probability (within 0.05) rather than log probability, which rounding moves for unlikely words. Measured on llama.cpp b11149, every catalog model tried reused 97–99% of a long chat, and --swa-full added nothing, so Toskar doesn't turn it on (#838).
  • The weekly quality run checks the prompt cache against a real model: how much of a chat's later turn comes from the cache (at least 80%), and that an answer from a warm cache matches a cold one, token for token, with top-5 log probabilities within 0.05. The report gives the tokens read and cached each turn. TOSKAR_QUALITY_RUN picks the tests scripts/quality-model.sh run runs (#838).
  • Context windows past 32K tokens use an 8-bit key/value cache, about half the memory, which answered as well in the quality set. Choosing 64K or 128K for a model, or an API request that needs it, now fits more computers, and the memory estimate counts the smaller cache (#962).
  • The quality sets fail only when their pass rate's interval says so: the whole 95% interval under the bar, or, given a baseline run with TOSKAR_QUALITY_BASELINE, a significant drop from it (p < 0.05). A rate just under the bar but within run-to-run noise passes with a note (#834).
  • The quality set can run each case more than once and say how sure its pass rate is. TOSKAR_QUALITY_REPEATS runs each case with a seed per run, and the report gives the pass rate with a 95% interval, the flaky cases, and, from three or more runs, pass@3 and majority of 3. TOSKAR_QUALITY_SAVE keeps every answer, and TOSKAR_QUALITY_RESCORE checks them again after a check changes, without a model. A report says when a run didn't cover every case. The weekly real-model run uses three repeats (#834).
  • Answers that repeat or edit text already in the chat come faster: chat models guess the next words from the chat and check several at once (llama.cpp's n-gram speculation), when the installed llama.cpp has it. On a 7B model on a Mac, editing pasted code was 42% faster and editing a pasted letter 99% faster, other answers 0 to 7% faster, and every answer was the same as before. It needs no extra model or memory. A run's details show how many guessed words were kept, and TOSKAR_SPECULATIVE=off turns it off (#835).
  • The user guide's code examples are tested. A code block with a test id runs in CI against a throwaway Core, with the repository's other examples, in sh (with the real curl and toskarctl), Python, TypeScript, or Go. Its output must match what the page shows, with … for what changes. A test id on a block CI can't run fails the user guide check (#882).
  • The user guide holds detailed pages. A section can add a summary, a kind (task, concept, or reference), and blocks: subheadings, steps, "What you need" lists, tips and warnings, tables, screenshots, code examples, and related sections. scripts/check-user-guide.py holds a detailed section to the bar in docs/user-guide/STYLE.md: a screenshot or a tested example, plain sentences, and no device brand names. A pull request that adds a feature changes the guide too, or has the no-user-docs label saying why (#882).
  • Install and start your first chat says what each amount of memory runs, adds the apt repository, and shows the download finishing and a first answer. Run without a desktop is a full page: the service, opening Toskar from your own computer through SSH, being the Owner, updating, and troubleshooting. The install script and apt lines on the page are the ones CI runs on a fresh Ubuntu computer: an example can name the test that runs it with runs_in (#885).
  • Chat covers what each model can read when you attach a picture, answers from the web and what they need, and which version of an answer a chat continues from, with a generated screenshot of an attached spreadsheet read, cited, and answered with a made spreadsheet (#886).
  • Choose and manage models explains what each kind of model is for, community ratings and exactly what sharing one sends, new models in the News, that a model needing more than the graphics card is tight, and toskarctl models. New generated screenshots show Browse all models with a download under way and Add a model › From another app (#887).
  • Use Toskar on your phone, tablet, TV, or watch covers chatting with a model on the device itself, when it's offered, and what leaves the device, and turning on notifications from your computer (#888).
  • The user guide's screenshots are captured from the web UI with demo data instead of by hand. A page names a screen from docs/user-guide/screens.json, and scripts/capture-docs-screenshots.sh saves it. CI captures them on every pull request that touches the interface, says which ones changed, and commits them on main, so the guide at a release shows that release's interface (#882).

Fixed

  • "Allow for session" on a tool's approval now lets that tool run for the rest of the chat without asking again. It was ignored for tools set to Ask, and where it did apply it covered every chat and person until Toskar restarted. A tool that asks because the chat read someone else's writing still asks each time. toskarctl chat offers it as [a]lways in this chat (#959).
  • /v1/chat/completions answered every failure while answering as `500 CHAT_FAILED` with the runtime's text. It now gives the stable code, as chat streams do, and a conversation too long for the model is a `400 CONTEXT_TOO_LONG` the caller can fix (#840).
  • A clip's length now comes from what you asked for. "Just one second long" made two seconds when the model left the length out or passed the default. Toskar reads the length from your message, such as "one second", "3s", "half a minute", "zwei Sekunden", or "3秒", and a clip asked for longer than clips can be is made as long as they can be, with a note saying so (#789).
  • Connect a device warned that over plain HTTP others on the network could see the code, even when the connection was encrypted and the dialog said so on the line above. The warning now shows only for a connection without the computer's certificate (#213).
  • A request to a model that wasn't running was checked against 8,192 tokens, though the model would start with a bigger window (#962).
  • When a model ended its turn without writing anything, the chat failed at once with "The model didn't write an answer". Toskar now asks the model once more, quietly, and shows that error only if it writes nothing again. It doesn't ask again after a tool has acted (#1001).
  • A model bigger than the graphics card runs, partly on the processor, instead of stopping while it loads. Toskar no longer asks llama-server to put every layer on the card; llama-server fits as many as the card's free memory holds, a draft's too, and keeps the rest in system memory, a mixture-of-experts model's experts first. An older llama-server without this still gets every layer on the card (#955).
  • With a draft model, Performance showed the draft's layers on the graphics card instead of the chat model's (#955).
  • A model's fit on the Models page now says how well it runs. A model that needs more than the graphics card's own memory runs partly on the processor, so it's tight at best and says so, where a 32B model was an excellent fit for an 8 GB card. On a computer without unified memory, some memory is kept for the operating system, so a 9B model is no longer a good fit with 8 GB. Best overall, Best quality, and the other badges go to models that fit well, and Best overall is the model setup recommends (#932).
  • Models like Qwen 3.5 and 3.8, which keep a full cache in only one layer in four, no longer have their memory overestimated. The catalog's full_attention_interval, or the model file's own layer list, counts the cache only for those layers, so the fit, the context gauge's memory, and the Auto window are right for them: a Qwen 3.8 27B gets a 32K window on a 24 GB card instead of 8K (#962).
  • A question about your own life, such as "Where does my sister live?", could be searched on the web when the model first said it didn't know, and answered from pages that had nothing to do with you. Toskar now asks the model once more to answer from the chat and its summary, and to say it doesn't know if it isn't there. Such a question is never looked up (#834).
  • When a long chat was summarized a second time, the new summary could cover only what was said since the first one, and lose everything before it. The summary now keeps the earlier summary's facts (#834).
  • The weekly prompt cache check no longer fails when a CPU reads the warm and cold prompts in different batches. Warm and cold must write the same tokens, each with the same probability within 0.2, until they part at a near-tie that batching can tip; gaps on tokens neither writes are reported, not failed on. Batching alone moves probabilities by up to 0.15 on a CPU and, at a near-tie, the word written (#838).
  • A real model's turn that failed, such as one where it wrote no answer, failed the whole quality run even when enough cases passed. The 1.11.2 release stopped this way at 64 of 70 (91%), over its 90% bar. It now counts as one failed case (#834).
  • A long chat could fail with "Context size has been exceeded" when two requests to its model ran at once, such as a turn and the chat's summary, or Deliberate's drafts. llama-server ran four requests sharing one window's room while telling each it had all of it. A model now takes one request at a time with the whole window, and the others wait their turn. The weekly runs caught it on Qwen 2.5 7B and 32B (#834).
  • A long chat's summary could be an answer to its last question instead of a summary, so it lost what was said earlier, such as where the person's sister lives. The summary request now asks for the summary after the conversation, and to keep what the person said about themselves and the people and places in their life (#834).
  • Claude Code's requests to /v1/messages were refused: a system turn partway through the conversation now becomes instructions, and the tools it adds join the request's. A raw request's window counts at most 8,192 tokens of its max_tokens, since agents ask for 32,000 as a ceiling, and a model with a 32K window was refused (#959).
  • On Windows with an NVIDIA card, where llama.cpp names the card with its PCI address, Performance and the model's GPU status showed the name with part of the address, such as "RTX 3080 Ti Laptop GPU) (0000:01:00.0" (#955).
  • On Linux, an AMD or Intel card's name was lspci's whole line, such as "03:00.0 VGA compatible controller: Advanced Micro Devices, Inc. [AMD/ATI] Navi 31 [Radeon RX 7900 XT/7900 XTX] (rev c8)". It's now the name people know, such as "AMD Radeon RX 7900 XT/7900 XTX" (#955).

Security

  • Text on a web page, in a file, in a Knowledge passage, or in a tool's result that gives orders to the AI is left out before the model reads it: forged chat turns written in a model's markers (<|im_start|>, [INST], and the like), and sentences such as "ignore previous instructions" or "note for AI assistants". A note stands where the text was. Labelling that text as data wasn't enough: a 7B model followed a page's instruction to end its answer with a word every time, and now never does. The sentence filter is English only, so a tool that changes something still asks first after an answer reads untrusted text (#938).

1.11.2

October 27, 2026

Downloads and release notes ↗Docs for 1.11.2 →

Added

  • A reasoning model can answer without thinking first, which is much faster for a question that doesn't need it: reasoning_effort: "none" on /v1/chat/completions (contract 1.60) and `toskarctl ask --no-thinking`. Toskar uses the model's own switch where its chat template has one, such as Qwen 3's, and otherwise a thinking budget of none, as for DeepSeek R1's distills, which always think. It also works on a paired computer, and a model that doesn't think is unaffected.

Changed

  • toskarctl ask and the Go examples use the official Go SDK 0.2.0, built for client contract 1.59.

1.11.1

October 26, 2026

Downloads and release notes ↗Docs for 1.11.1 →

Added

  • examples/ci has workflows that ask Toskar as a CI step: label new issues (GitHub Actions and Forgejo) and explain a failed build from its log (GitHub Actions and GitLab CI). Each one keeps the issue or log as data, and changes only what it allows, in a separate step. Their shell steps run in CI against a throwaway Core, as the other examples do.
  • A triage set in the quality tests: 30 of Toskar's own issues with the labels maintainers gave them. It measures how well a model suggests them the way the Triage workflow asks (#455), as precision, recall, and how often the issue's type is right, and it runs in the weekly real-model run. toskarctl ask's fencing of data moved to internal/askdata, so the test sends exactly what the workflow does.

Changed

  • A request to /v1/chat/completions that names a model gets that model, or an error saying why. When the model couldn't answer, such as one that failed to load, another installed model used to answer in its place without saying so, which made results that compare or measure models wrong. auto and profiles still fall back, and `"toskar": {"fallback": true} allows it for a named model too (contract 1.59). toskarctl ask --model` gets the model it names.

Fixed

  • Install in Browse all models now shows what happens. It said Install again as soon as the request was sent, while the model downloaded in the background, so it looked as if nothing happened. A row now shows Starting…, then Downloading… with how far along it is, then Installed, as the catalog's cards do. A download that fails says why and offers Install again.
  • Structured output holds on an external OpenAI-compatible server and on a paired computer too. The JSON Schema reached only this computer's llama.cpp, so a model elsewhere could answer outside it, and the request failed with 422 INVALID_STRUCTURED_OUTPUT. An external server now gets it as response_format (and is asked again without it if it refuses), and a paired computer's llama.cpp constrains the reply as this one's does.
  • A model loading for a request is no longer unloaded as idle partway through. The idle unload went by the model's last use, which was the previous instance's, often long ago. So a model that had been unloaded once could be stopped again while it loaded, and the request moved to another model or waited a long time. A load in progress now counts as use, and the model stays loaded until it finishes.

1.11.0

October 24, 2026

Downloads and release notes ↗Docs for 1.11.0 →

Added

  • News: new open models, runtime releases, and AI labs' announcements, a few lines each, from the AI news toskar.ai publishes every morning. A model in Toskar's catalog shows how well it fits this computer, and Admins can install it from there. Toskar reads the news only when News is opened, at most every six hours, keeps a copy for offline, and lists each read in What left this computer; Settings › AI news turns it off.
  • Text for Ask Toskar on selected text in the phone app (yeixio/toskar-apps#276): rewrite, reply, translate, explain, summarize, and fix spelling and grammar, with each action's instruction to the model, in every language.
  • For a CI job on the API Access page makes a key that only asks: it reaches the OpenAI-compatible chat, with no memories, knowledge, or tools, so what a job sends, such as an issue's text or a build log, can't reach your data or make anything happen. Anything else answers 403 KEY_CHAT_ONLY, MCP included. A key's permissions can also be set to chat only later, and POST /api/v1/api-keys takes "preset": "ci" or permissions from the start (contract 1.57).
  • docs/connecting.md says exactly how an app or SDK finds a computer, pairs with it, pins its certificate, and reaches it from anywhere, with test vectors in api/connection-vectors.json that Core's tests keep in step with its own code.
  • The outside-network check also runs the official SDKs' conformance programs, as devices that pair at home and connect from away, both directly and through the relay.
  • Release notes thank everyone outside the maintainers who has a commit in the release.
  • GOVERNANCE.md describes the contributor, triager, reviewer, and maintainer roles, and MAINTAINERS.md lists who holds them.
  • install.sh ci runs Toskar for a CI job on the runner itself: it installs the release in a folder, with no root or service, starts it, installs a model, and sets TOSKAR_URL and toskarctl for the job's next steps, in GitHub Actions, GitLab, Forgejo, or anywhere with a shell. Toskar in CI puts it together with toskarctl ask and a key for a CI job.
  • Releases include Linux headless archives (amd64 and arm64), the same programs as the packages, for installs without root.
  • Toskar Pro's 14-day trial and Unlock (yeixio/toskar-apps#336): the Pro desktop app from the Mac App Store or the Microsoft Store tells Core where its purchase stands, GET /api/v1/license reports it, and once a trial has ended without the Unlock, installing models, creating automations, and adding people wait for it (UNLOCK_REQUIRED). Everything already set up keeps working. Core on its own is unaffected.
  • On the Mac App Store, Toskar Pro's free 14-day trial is started by the person, as Apple asks: until then the banner says how long the trial lasts and what the Unlock costs, with Start free trial, and new setup answers TRIAL_NOT_STARTED (yeixio/toskar-apps#336).
  • toskarctl ask sends one prompt and prints the answer, for scripts and CI jobs: gh issue view 12 | toskarctl ask "Summarize this bug". Piped text is sent as data apart from the prompt, --json-schema answers with JSON that fits a schema, and memory, knowledge, and tools are off unless asked for. It's built on the official Go SDK.
  • The TV app's demo, for trying it without a computer, in every language.

Changed

  • The examples use the official SDKs, in Python, JavaScript, and Go: streaming an answer, watching events, and connecting as a device at home or from anywhere. The curl examples stream through the OpenAI-compatible API and read TOSKAR_URL and TOSKAR_API_KEY.
  • Every example runs in CI against a throwaway Core, on each pull request and each night, so one that breaks fails the build. The Python events example prints each event as it comes, even into a pipe.
  • The examples install the SDKs from npm and PyPI (toskar 0.1.0), and the Go examples use the Go SDK's v0.1.0 release.
  • The phone and TV setup steps name Toskar Core first, and open Toskar at http://localhost:7331, which works with Core and Pro alike.

Fixed

  • Clips from Wan 2.2 TI2V 5B, the default clip model, no longer wash out toward white as they get longer, and the shortest no longer come out as noise. It now samples 30 steps with guidance 5 and shift 5, so a clip takes about 1.5 times as long as before.
  • docs/connecting.md says to send the device's key with the health check, since a computer that asks for keys answers it only with one, and to check the certificate on resumed TLS sessions too.
  • A model file put in Toskar's models folder by hand now shows up: the Models page says it isn't added yet, and Add model › From another app adds it where it is, without a copy.
  • Pictures sent to /v1/chat/completions reach the model. They were read, then dropped, so a model answered without having seen them. They're now shown to a model that can see, as pictures attached in the chat are, and with none, the answer says so. A picture is a data: URL, or an http(s) one Toskar fetches from the internet only.
  • Answers from /v1/chat/completions look like OpenAI's to the libraries and tools that read them. Each has its own id and a created time, usage counts the tokens (cached and reasoning ones too, and a stream sends it with stream_options.include_usage), and finish_reason says length when an answer was cut short at its token limit, instead of always stop. Errors use OpenAI's shape, type, param, and code, including an error partway through a stream.
  • The API description (api/openapi.yaml) gives the client contract's version (1.55, not 1.4.0), and no longer lists one response twice.
  • The API description now describes people, sign-in, invites, roles and profile pins fully, with named types, including fields it left out (an API key's pinned profile, and a person's email and portal).
  • It describes connecting a device, too: the pairing code, the paired device's key and certificate, and its route for access from anywhere.
  • It describes every type in the client contract (messages and their sources, steps, files and deliberation, run traces, events, notifications, and errors) from the code's own fields and comments, and the routes for chats, their summaries, and versions use them.
  • It describes the chat stream's events (token, done, error_code, and error) and the events stream, so a client can read each event as a type. Chat answers' metrics include cached_tokens (contract 1.56).
  • It describes models, runtimes, and ratings fully: the model list, recommendations, browsing, installing, importing, running models, and where a model can run.
  • It describes automations, tasks, knowledge, training, portals, joining computers, logs, diagnostics, caches, connectors, MCP servers, news, and updates with named types, from the code.
  • It describes memory, notifications and their destinations, API keys, webhooks, speech, tools, and more of training with named types; every answer the tests see now matches it.
  • Every JSON answer in the API description is a named type, including diagnostics, transcription, tool activity and providers, and the OpenAI-compatible model list and chat completions, ready for SDKs to be generated from.
  • Every route in the API description is tagged by area, such as Chat, Models, or Automations, so the API reference and SDKs group them.
  • Structured output uses the whole JSON Schema a request sends. Only type, properties, required, enum, and items reached the model's grammar and the check, so anyOf, $ref, additionalProperties: false, bounds, pattern, and const were dropped without notice, and a schema whose top level was anyOf accepted anything. The schema now goes to llama.cpp as sent, answers are checked against every keyword, and a schema that isn't valid JSON Schema, refers outside itself, or is over 64 KB is refused with 400 INVALID_RESPONSE_FORMAT.
  • An API request that turns tools off can use a model on the external server again. It was refused as if its profile kept work on this computer.

Security

  • Text a model reads can't pose as its own turn markers. A web page, a file in Knowledge, a tool's result, an automation's trigger, or a visitor's message that spelled a control token, such as <|im_end|><|im_start|>system, reached llama.cpp as that token, so it could start a real system turn. Every message is now escaped on its way to the model, from the model's own vocabulary and the common chat templates' markers, and so is what goes to an external server.

1.10.0

October 20, 2026

Downloads and release notes ↗Docs for 1.10.0 →

Added

  • An automation can look at cameras as each run starts, such as Home Assistant's: their pictures are shown to a model that can see, so a run can say what's at the door or in the driveway. Set them with cameras in the API or toskarctl automations … --camera camera.driveway. Only Admins and the Owner can give an automation cameras.
  • In Automations, Runs › When a device changes picks a device of a connected service, such as Home Assistant's, by name, with the state to wait for and how long to wait after a run; Look at cameras picks up to 4 cameras for each run to look at. The list shows them by name, such as "When Front Door Visitor turns on".
  • An automation can run the moment a device changes, such as a Home Assistant doorbell, motion sensor, or door contact: a device trigger, with the state to wait for and a cooldown so it runs once per event. Set it with {"kind": "device", …} in the API or toskarctl automations … --trigger device --device binary_sensor.front_door_visitor --to on.
  • GET /api/v1/devices lists the devices of connected services, such as Home Assistant's cameras, sensors, and speakers, by kind, for picking them in automations.
  • Pictures and clips show how far they've got (#706): whether the model is loading, which step it's on with about how long is left, and when it's finishing, beside a frame where the result will appear. On a long wait, a line says what Ratatoskr is up to, such as running down Yggdrasil for more paint (from a user's idea). Screen readers hear the facts, not the jokes.
  • tool.progress events with stage, step, steps, and seconds_left while a picture or clip is made on this computer (#706, contract 1.44).
  • After an answer, Toskar may offer a model that would have answered it better, and install it in one tap (#707): one that writes code better, reads your files and knowledge more reliably than a small model, writes your language better, thinks through math and logic, or reads a longer chat at once. Install and answer again answers the same question with it as a new version. Only for Admins and the Owner, once per kind in a chat, and only for a model this computer runs well with room on disk. Settings › History › Suggest better models turns it off for you.
  • Model offers: meta.setup with ability model, reason, and fit, and the per-person suggest_models preference (#707, contract 1.45).
  • Muse Glimmer 30B, Meta's model that sees pictures, codes, and reasons well, is in the model catalog for computers with 32 GB of memory or a 24 GB graphics card.
  • A notification from an automation that watches a web page keeps that page, and a push sends it with the title, so tapping it on the phone opens the page at once, even away from home. A private push leaves it out.

Changed

  • Research recommends Muse Glimmer 30B on computers with room for it, about 24 GB of graphics memory or a 32 GB Mac. It answered every reasoning question right on an RX 7900 XTX, matching Qwen 2.5 32B, and also reads pictures.
  • FLUX.2 [klein] 4B, Z-Image-Turbo and Wan 2.1 T2V 1.3B now also work out their graphics memory from parts measured on Apple silicon, as FLUX.2 [klein] 4B high quality and Wan 2.2 TI2V 5B do. Wan 2.1 T2V 1.3B needs more than its old estimate, up to about 15 GB of graphics memory for a short clip.
  • The graphics memory FLUX.2 [klein] 4B high quality and Wan 2.2 TI2V 5B need is worked out from measured parts and the requested size and length, instead of one generous estimate: about 6 GB for a 1024×1024 picture and 7 GB for a 2-second clip, rather than 12 GB. Smaller graphics cards get "Works well" where they can.
  • A picture or clip too big for this computer's graphics is turned down before it starts, with the size or length that fits.

Fixed

  • An automation created through the API without a profile, model, or time zone now gets General Assistant, Auto, and this computer's time zone, instead of being saved and then failing every run. A profile that doesn't exist is refused when it's saved. toskarctl automations create --name no longer needs --profile and --model.
  • An automation that looks at a camera answers from the camera's picture: it no longer looked up the camera in Home Assistant and answered with its status, such as "the person is currently idle".
  • Auto treats requests like "Write a Swift function…", "a string in Go" or "this Rust trait" as coding, so it picks a coding model and suggests one when none is installed.
  • A phone, tablet, or TV that removes this computer now disconnects from it, with its watch. Before, the computer kept the device connected and kept sending it notifications. The apps use the new DELETE /api/v1/me/device.
  • A chat's stream ends once the answer is saved, so the phone app no longer misses an answer it reads back as the stream ends, as when the first answer in a new chat suggested a better model.
  • On a Mac, how well a model fits now counts the part of memory the graphics may use, and a model's picture projector, so a 24 GB Mac is no longer told a 30B model runs comfortably. Recommendations and better-model suggestions follow it.
  • The web app now shows "Reasoning…" while a model thinks (#701) and "Compacting earlier messages…" before a long turn (#342): it never subscribed to those events, so neither line appeared. A test now checks that the web subscribes to every event the chat page handles.
  • What left this computer no longer lists a notification push that no phone or tablet took.
  • Disconnecting a device, from Devices, API Access, or by removing a person, now also removes its notification registration; ones left by devices disconnected before are cleared.
  • Run now on an automation that watches a page, feed, or folder now looks at what it watches. Before, the run wasn't told the page, so it answered without it.
  • The first-run setup no longer offers to download a model that's already here, and its estimated active memory is what the recommended models need, not the computer's whole graphics memory.
  • A model file added from another Toskar, or kept under its source's name as LM Studio does, is recognized as that catalog model: it's installed under the catalog's name, with its details, instead of as a new model.
  • The Automations page no longer fails to open after an automation is made without a time zone, through the API, toskarctl, or a request the computer read. Toskar now fills in this computer's zone by its name, such as America/Juneau, and reads ones saved before as that.
  • When a model calls a tool Toskar didn't offer, as Muse Glimmer 30B sometimes does, Toskar asks it once more, saying which tools there are, instead of failing with no answer.

1.9.2

October 18, 2026

Downloads and release notes ↗Docs for 1.9.2 →

Added

  • A reasoning model's thinking is no longer a silent wait (#701). The chat says Reasoning… while it thinks, the answer's steps say how long it thought ("Thought for 12 s before answering"), and Thinking under the answer opens what it thought. The model's thinking was thrown away before.
  • chat.reasoning events (started, then done with ms), the answer's meta.reasoning (ms and text), and include_reasoning on the OpenAI-compatible API, which returns the thinking as reasoning_content (#701, contract 1.43). External servers that send reasoning or reasoning_content work too.
  • GET /api/v1/settings tells Admins who could connect from another computer, in network_access: people, devices, keys, and whether a sign-in service is set up (#711, contract 1.42).

Fixed

  • Memory commands work in all ten languages, not only English (#718): „Merk dir, dass …“, « Souviens-toi que … », „…, merk dir das“, 「…を覚えておいて」, “记住:…”, and the forget and “what do you remember about me” forms. The help text already told people to say them in their language, but only English was recognized, so a small model often said it would remember and saved nothing. Each language's words are in i18n/memory/.
  • Toskar on phones and tablets no longer shows 57 messages in English in the other nine languages (#716): memory confirmations, the context gauge, the memory and web lookup settings, and the steps, notices, and errors of chats on the device. A catalog check now fails when a language misses a key English has, unless the language is marked partial.
  • Setting up people no longer looks like it needs an API key (#711). Turning on local network access asked for a key, even though people you add sign in with their username and password. API Access now says who connects how (people with a password, phones and tablets with a code, apps and scripts with a key) and lets you turn it on once anyone could connect. People offers Turn on network access right under a new person's link. The user guide's People section has the steps in the order that works.
  • Sunrise and sunset questions are looked up instead of guessed (yeixio/toskar-apps#255). "What time is sunrise in Chicago this morning?" didn't count as needing current information, so a small model made a time up (6:36 AM, that day's dawn). Sunrise, sunset, dawn, dusk, first light, and golden hour now count, and the answer reads today's times for the place from the weather service, in its local time, before any search result.

1.9.1

October 18, 2026

Downloads and release notes ↗Docs for 1.9.1 →

Added

  • User guide: Get more careful answers, covering Effort, reasoning models and why Auto doesn't pick one for a quick question, a profile's Reasoning level, comparing independent drafts, and what each costs (#697). The Chat section and the API's reasoning_effort point to it.
  • A step that takes a while, such as making a picture or a clip, shows how long it has taken beside it after 15 seconds, so a long wait doesn't look stuck (#705).
  • Toskar on phones and tablets has the words to say a picture or clip is being made on your computer, in the 10 languages (#705, yeixio/toskar-apps#290).

Changed

  • A computer tells Toskar's relay its Toskar version when it connects, beside its name, so the relay's operator sees which version each computer runs on the relay's dashboard (toskar-relay#47). It goes only to the relay, never to devices or into records.

Fixed

  • Asking to make a picture ("make a picture of a cat riding a bike") no longer offers to take a photo instead: the typo fixer read "make" as a misspelled "take". It now leaves 131 more everyday words alone, such as make, think, show, short, and plan (yeixio/toskar-apps#195).
  • Phone text for subscribing to access from anywhere when the store hasn't passed a renewal on yet, and for Restore purchase finding nothing, in all ten languages (yeixio/toskar-apps#285).
  • When access from anywhere ends because its token ran out, as a lapsed subscription's does, Settings › API Access says so, "Access from anywhere ended on Oct 15, 2026", with Subscribe again, instead of "Can't connect to relay.toskar.ai (TOKEN_EXPIRED)". Toskar stops sending the expired token to the relay every few minutes, logs once, and a new token from the Toskar app or a subscription takes over within a minute, without a restart (#692, contract 1.41: relay state ended, and subscription ended with expires).

1.9.0

October 17, 2026

Downloads and release notes ↗Docs for 1.9.0 →

Added

  • API Access sells Toskar's relay itself: Subscribe for $4.99 a month or $39.99 a year opens the checkout in your browser, access from anywhere turns on a few seconds after you pay, and Manage billing changes the card or cancels. A subscription bought in the Toskar app shows as such, and is managed in the device's store (#639).
  • Access from anywhere can be bought on the computer itself, without a phone: POST /api/v1/remote-access/subscription/checkout opens a Stripe checkout (monthly or yearly), Toskar turns access from anywhere on once it's paid and renews it as the subscription does, and POST /api/v1/remote-access/subscription/manage opens the billing page for changing the card or cancelling. GET /api/v1/remote-access says how the subscription stands (#639, contract 1.38).
  • The chat shows auto-compact (#342):
  • The context gauge explains it and offers View summary, Compact now, New chat from summary, and Keep this chat verbatim.
  • A divider in the chat marks where the summary takes over, and opens it to read and correct.
  • "Compacting earlier messages…" shows while a long turn compacts first.
  • Settings › Chats & history chooses when it starts.
  • Auto-compact has a name and controls (#342):
  • A long chat's older messages are sent to the model as a summary, and every message stays saved. Now you can read the summary and correct it, compact a chat now, start a new chat from its summary, or keep a chat verbatim.
  • Settings › Chat chooses when it starts: Early, Normal, or Late.
  • A turn that wouldn't fit compacts first, instead of leaving the oldest messages out.
  • API: GET and PUT /api/v1/conversations/{id}/summary, `POST …/compact, POST …/from-summary, a chat's compact_off`, and the compact_at setting (contract 1.35).
  • Administer → People has What people may use: a switch for each capability and role, and under each person their own Allow or Block (#589). The app hides what's off for the person using it: attaching files, the model picker, read aloud, and making automations.
  • Admins choose what each role, and each person, may use (#589):
  • the capabilities are files, pictures in, making pictures, making video, voice, the web, Knowledge, adding to Knowledge, automations, tools that change things, choosing a model, and access from anywhere;
  • Visitors start without files, making pictures or video, or automations, and Members without tools that change things;
  • Toskar enforces them everywhere a person comes in, and `GET /api/v1/me` lists them so apps can hide what's off;
  • asking for something turned off gets a plain answer saying who can turn it on. See What people may use.
  • A connected phone can connect its watch to the computer, from anywhere and with no code to type: POST /api/v1/me/devices (contract 1.33) gives the watch a key of its own, for the same person, so it shows in Devices and can be disconnected on its own (yeixio/toskar-apps#213). A person may have 20 connected devices.
  • User guide: Report a problem, covering what a report sends, what's taken out first, how the TV and watch send theirs, and that Toskar sends nothing on its own, so there's nothing to turn off (#588).
  • The media quality set asks its judge whether a picture or any frame of a clip is garbled or corrupted, or flashes or jumps apart from its neighbors, so a broken decode fails even when it shows the right thing; and a Mac-only clip case guards Apple's video decode path (#553).
  • User guide: What pictures and clips need, with each model's memory, download, and measured speed, the verdicts, what happens when memory is short, and the Mac App Store limit (#552). The Media setup card and a chat's offer to set one up when it may be slow or tight link to it, for the guide of the running release.
  • toskarctl people: list, add, show, change, link, remove, and see the keys, devices, and activity of the people who use this Toskar, from a script or a server without a browser (#592). Passwords come from standard input only.
  • TOSKAR_ADMIN_EMAIL and TOSKAR_ADMIN_PASSWORD_FILE add an Admin when a new server starts, once, for Docker and Kubernetes installs set up from a script.
  • Manage people without a browser (#592). Through the API, Admins can:
  • add someone with an email address and a password they can sign in with at once, which they can be made to change first;
  • set a password for someone;
  • look up one person, or filter people by role or by disabled;
  • remove someone, deleting their chats, memories, automations, and files, or giving them to another person;
  • see and revoke someone's API keys and devices. Every change is logged in GET /api/v1/people/activity. Someone who must change their password is asked for a new one when they sign in.
  • Core's side of notifications in the Toskar app (#460): an app notification destination that pushes to the person's paired phones and tablets through Toskar's push relay, with the title only (the default), the full text, or a private notice, and GET, PUT, and DELETE /api/v1/me/push for a paired device to register the sealed handle the relay gives it. Each push is signed with the computer's key; the computer never sees a device's push token. It takes effect once the relay and the apps support it (docs/push.md).
  • The first phone or tablet that turns on notifications in the Toskar app adds the Toskar app as a notification destination, with the title only, so turning them on in the app is all it takes (#460). The app's text for notifications is in the 10 languages.
  • Settings › Notifications can send to the Toskar app on your phones and tablets (#460): Add the Toskar app, with the title only (the default), the title and text, or a private notice, and how many of your devices have notifications on. Your devices shows which ones do. Pushes arrive once the apps and Toskar's push relay support them.
  • POST /api/v1/diagnostics/report sends a problem report to Toskar's report service, which files it as an issue grouped with others for the same error and version, instead of the person writing an email. It's listed in What left this computer, an Admin can turn it off with the problem_reports setting, and paired devices may send through it too (#588, contract 1.40).
  • Report this, beside errors in chat, on a page that couldn't load or ran into a problem, in automation runs, and when a model won't start or install (#588). It shows everything the report sends: what went wrong and where, the Toskar version, the kind of computer, and for an Admin Toskar's last warnings and errors. You can add what you were doing. Email addresses, keys, home folder names, the computer's name, and internet addresses are taken out first. Send by email opens your mail app addressed to info@yeix.io, and Copy is there if email isn't set up.
  • GET /api/v1/diagnostics/report (contract 1.32) gives that report, scrubbed. Members may call it; only an Admin's report has the log.

Changed

  • The Mac App Store's copy of Toskar doesn't sell access from anywhere: API Access says to subscribe in the Toskar app on a phone or tablet, as the App Store requires for an app it sells, and a subscription bought there covers the computer as before (#630).
  • Search boxes forgive typos: "mistrl" finds Mistral in Models, and "atomations" finds a chat or automation named Automations (#451). Exact matches still come first, and searches of three letters or fewer stay exact. This covers the chat list, Models (both tabs), Tools, and Automations.
  • Report this is now beside the errors the web app shows when an action fails or a page part can't load, about 90 in all, from saving settings to training and connecting a computer (#588). Each shows its text in the App language from its code and sends the code with the report. Sign-in and password errors, form checks, and the public chat portal stay as they were.
  • When a picture or video model may be slow or tight, the Tools page and the chat's setup offer say what it needs and what the computer has, such as "Needs about 16 GB of graphics memory; this computer's graphics card has 10 GB" (#550).
  • On Apple silicon, models are checked against the share of memory the graphics may use, about two thirds, or three quarters above 36 GB, so a model that needs more is marked as possibly slow before it's set up.
  • The chat's setup offer carries the model's verdict and these numbers (verdict, short, need_bytes, have_bytes, contract 1.34), and GET /api/v1/images/setup and /video/setup add them to each model, with the computer's graphics_bytes and shared_graphics.
  • The errors that lists show now carry a code beside their text (#588, contract 1.36): automation runs, previews, and last errors, knowledge sources, tasks, and failed model downloads. The web app shows them in the App language and sends the code with Report this. A failure without a code of its own gets one for where it happened, such as AUTOMATION_FAILED or KNOWLEDGE_SYNC_FAILED, and a run or task cut short when Toskar stopped is AUTOMATION_INTERRUPTED or TASK_INTERRUPTED.
  • An error on another computer now reaches the app with its own code instead of as plain text (#588). Paired computers answer each other with coded errors, such as MODEL_NOT_INSTALLED from the computer running a model, and PEER_UNAUTHORIZED or PEER_UNAVAILABLE when one won't take a request. Computers on an older Toskar still get, and send, plain text.
  • Report this sends the report to the Toskar team from the app and the desktop app, where it becomes an issue, instead of opening your email app: it says when others hit the same error, Include details chooses whether the computer's part goes, and an email for a reply reaches only the team. It falls back to email when it can't be sent, and an Admin can turn sending off in Settings, under Problem reports (#588).
  • Automation requests with typos are understood too: "evry mornign at 8 chek the price" runs every morning at 8, and "every wenesday" on Wednesdays (#451). Words typed without their accents count in every language written with spaces, so "todos los dias" is "todos los días".
  • Toskar understands typos in the requests it handles itself: "whats teh wether tomorow" looks up the weather, "remmeber that i like tea" saves a memory, and "helo" or "thnaks" is small talk (#451). It corrects a word only toward the words those rules look for, and leaves real words such as "whether", "plain", or "test" alone.

Fixed

  • On Windows, the desktop app's title bar follows Toskar's theme instead of Windows' own, so a dark Toskar no longer sits under a white title bar (needs the desktop app from yeixio/toskar-apps#251).
  • A picture or clip a tool made is no longer lost when the model then writes nothing: the answer keeps the file, with "Here's what I made." A turn where the model writes nothing at all now says so (EMPTY_ANSWER) instead of ending without an answer (#553).
  • Every error now carries a code for Report this (#588). Chat errors that had none send CHAT_FAILED, an error with its own code keeps it in the chat stream, and an API path no route takes answers ROUTE_NOT_FOUND or METHOD_NOT_ALLOWED instead of the web page.
  • A reply cut off by a lost connection now says so instead of looking finished. Every chat stream ends with done or error, including a stopped one.
  • The web app explains an answer it can't read, such as a proxy's error page, instead of showing a parse error, and no longer reports a cancelled request as the service being unreachable.
  • The media quality set's judge also fails a washed-out picture or clip, one faded until most of it is near white, which passed before (#553).
  • A quality run narrowed with -run counts only the cases it ran, instead of failing the ones it left out as stopped.
  • Pages no longer show defaults or "none yet" when their data couldn't be read (#217). Settings, API Access, Personalization, and network discovery wait for their settings before anything can be saved, so a failed read can't overwrite them. Chat history, a chat's messages, People, portals, keys and devices, what left this computer, the Media tab, benchmarks, training examples and plans, and installed and running models say they couldn't load, with Try again. An invite or public chat page that couldn't be checked says so instead of calling the link gone.
  • Diagnostics no longer shows every row as not responding before its first check, the empty Installed tab and Discover offer their next step, and Tools says when no tool matches the filter.
  • The smaller panels and lists say when they couldn't load, with Try again, instead of saying there's nothing or hiding (#217): your devices, chat cleanup, notification destinations and quiet hours, connected services, tool sources and their log, a tool's recent calls, the add-a-source gallery, models found in other apps, Browse all, base models, chats to train from, sample files, exports, the knowledge picker, topic attempts, computers' models, the external server, Connect a device, roles and profile pins, portal visitors and conversations, remote access, Diagnostics logs and tool activity, running models, recent benchmarks, the notification bell, and onboarding's hardware check.
  • Language, quiet-hours, and other settings controls wait for their saved values, and a choice whose list couldn't be read still shows in its menu. Browse all says when nothing matches, and the empty Activity page links to Chat.
  • On Portals, the link to turn on local network access is underlined, so it stands out from the warning around it by more than its colour.
  • The first chat after a paired computer connects no longer says a tool isn't set up, such as image generation, when the paired computer has it ready. Toskar waits briefly to hear what that computer can run, the first time only (yeixio/toskar-apps#195).

1.8.3

October 14, 2026

Downloads and release notes ↗Docs for 1.8.3 →

Added

  • Dev container configuration (.devcontainer/devcontainer.json) for VS Code and GitHub Codespaces, with Go 1.26, Node 22, and pnpm 9.15.0 pre-installed and port 7331 forwarded.
  • An Admin's paired phone or tablet can read whether access from anywhere works, and run Check now, for the summary in the app's Settings (#546). A Member's device still can't.
  • Diagnostics shows how fast answers start: the wait for the first word over the last week, for each app, with the model already running and with one loading first (#498). Chat requests can name their app in client, and run traces keep it and whether the request came from outside the home network. GET /api/v1/diagnostics/first-answers has the numbers.
  • A picture or clip a paired computer made says which one, beside its model, size, and seed.
  • Toskar notices another copy of itself running on the same computer, such as the installed service and a copy started by hand (#602). Access from anywhere names it, with its port, data folder, and version, since both copies connect to the relay as separate computers. Installing or upgrading the Linux package says so too.
  • An Admin can set up what a chat asked for, such as picture making or a model that sees pictures, from their phone: the answer shows what it would install, its size and computer, and its progress, and the chat finishes the request once it's ready. People who aren't Admins see who to ask. In all ten languages (yeixio/toskar-apps#198).
  • Access from anywhere has Copy details for a report: what it's doing, as text to paste into a bug report or support question (#602). It contains no tokens or secrets, and your internet addresses are replaced by their kind.
  • When the router won't open the port, access from anywhere offers **Try another port. When the relay already works, it offers Use the relay only**, which stops asking the router (#602).
  • Access from anywhere says how your router opened the port ("Your router opened port 7333 for this computer (UPnP)"), or why it couldn't, with the next step (#602). Possible reasons: no router answered, it refused, the port is already forwarded to another device, double NAT, carrier-grade NAT, or automatic opening is off. When the relay already works, the line is shown quietly.

Changed

  • A short question that needs the web or this computer, and any question from the watch, is answered by a model that is already loaded instead of waiting for a better one to load (#498). What I did says so when it skipped a load. Thorough effort keeps the usual choice, and automations are unchanged.
  • Settings has Keep models ready (15 minutes, 1 hour, 4 hours, or Until I stop them) in place of a minutes field. A model your phone, tablet, watch, or TV used in the last hour stays loaded either way, so the next question from it starts at once (#498).
  • FLUX.2 [klein] 9B is tested on real hardware: about 36 seconds a picture on a fast graphics card, within 32 GB of memory.
  • From a link no longer starts a download right away (#611, reported by @innovagent). Checking the link first shows what it would download: the size, the space left, and that it's a direct download saved on this computer, which Toskar doesn't charge for. When LM Studio, Ollama, llama.cpp, or GPT4All already has the same file, it offers **Use it where it is**: no download and no extra space. It won't start a second copy that doesn't fit.
  • From another app comes first in Add a model when another app has models Toskar could use.
  • While a model loads for an answer, chat says which model, on which computer, and about how long it took last time ("Loading Qwen 2.5 7B on Office, about 20 s"), and no longer says "Loading" when the model is already running (#498). A load on a paired computer now counts as load time in run traces and in Diagnostics' first-answer times.
  • When this computer connects to a relay, it sends its name, so the relay operator's dashboard can tell computers apart (#602). Only the operator sees it. It never reaches devices or the address record, and the relay still can't read anything inside a connection.
  • When only the relay reaches this computer from outside, Settings says "Reachable from outside your network, through the relay" in green. A direct way in that failed is shown quietly, as an optional speed-up, instead of in red as a problem to fix (#602).
  • When llama.cpp can't be installed, the error says which build it looked for, which releases it checked, and GitHub's HTTP status (with a hint when it's GitHub's rate limit), and where to put llama-server by hand (#64). docs/runtimes.md has the steps.
  • Model health events carry the end of what the process wrote (stderr_tail) and the last health probe's error (last_probe), so a diagnostics subscriber can show why a model stopped (#9).
  • Wan 2.2 T2V A14B is tested on real hardware, and asks for 48 GB of memory, as measured, instead of 32 GB: a 2-second clip took about 13 minutes on a fast graphics card. Below 48 GB it's set up only when you agree.
  • Wan 2.1 T2V 1.3B is tested on real hardware: about 3 minutes for a 2-second clip on a fast graphics card.

Fixed

  • On Computers, the header button for the open panel is the highlighted one: Add by command and Connect a device show which one is open, one at a time, and Find computers is the main button only while neither is open.
  • Asked to run a command that deletes or wipes things, the assistant no longer refuses and hands you the command to run yourself, which skipped the approval step (#557). It asks to run it with your approval when it can, or says plainly that it wasn't done; a command left in the answer is taken out, with a notice. Questions about how to do it still get the command.
  • The access-from-anywhere check says what answered instead of "Something other than Toskar answered on this port": another Toskar by name ("Toskar on linix-rig"), a web server that isn't Toskar, or this computer answering with an unexpected HTTP status (#602). The details need relay 0.3.4.
  • "Checked moments ago" is a note about the latest result, not an error. Settings shows the start of this computer's route ID, as the relay's dashboard lists it.
  • The route secret is made only once, even when several requests need it at the same moment. Before, a computer could enroll on the relay twice, under two routes, with its devices knowing only one (#602).
  • Access from anywhere says when your router has no internet address of its own, behind your provider's shared address (carrier-grade NAT) or another router, instead of that it refused to open a port, and suggests the relay.
  • Start on the Models page no longer does nothing when llama.cpp isn't set up, such as after adding a model From a link without going through onboarding (#610, reported by @innovagent). Start now sets llama.cpp up first and shows "Starting…" while it works. If it can't, it says why under the model, and the first chat that needs a model does the same. Every start attempt and how it ended is now in the daemon log.

Security

  • Toskar is built with the newest Go patch release and golang.org/x/net 0.60.0, which fix ten security problems govulncheck found in Go's web server, TLS, and HTML template code (GO-2026-6603 to GO-2026-6617).
  • Toskar's apt repository is signed, and the install steps no longer add it with [trusted=yes], which turned off apt's signature check (#599). New installs add its key from https://toskar.ai/apt/toskar.asc (fingerprint 7124 28C8 E821 E64C D418 32F5 2D28 4623 90D0 C604), and the source is now /etc/apt/sources.list.d/toskar.list. An existing install moves to the signed form by itself when the package upgrades.

1.8.2

October 12, 2026

Downloads and release notes ↗Docs for 1.8.2 →

Added

  • Phones and tablets connected to this computer can attach a photo, file, or recording to a chat, at home and away from home.
  • When a picture or clip has to be shorter or smaller than asked, a note under the answer says so, and why. For example: "Clips can be up to 5 s long for now, so this one is 5 s instead of the 10 s asked for."
  • On the Tools page, a picture or clip model that would be slow or too large here names a paired computer it works well on, with a button to set it up there.
  • Pictures and clips Toskar makes keep how they were made: the model, quality, seed, size, and prompt. The chat shows the model, size, and seed under each one, with the prompt on hover, so it can be made again.
  • A profile can choose the quality its pictures and clips are made at when a request doesn't ask for one: fast, balanced, or best (Profiles & Orchestration → Behavior).
  • An answer's steps say which model made or edited the picture or clip.
  • Pictures and clips come in three qualities: fast, balanced, and best. Asking for "a high-quality picture" or "a quick draft" uses the set-up model for that quality, and when it isn't set up, the answer says which model would give it.
  • Each picture and clip model says whether it works well on this computer, may be slow or tight, or isn't supported, counting a graphics card's own memory as well as system memory. A model that would be slow or tight is set up only when you agree.
  • Picture models Toskar can't give out can be set up with your own license: accept the license on Hugging Face, add your Hugging Face token on the Tools page, and confirm your license covers how you'll use it. The files come straight from the vendor. The first is FLUX.2 [klein] 9B, under the FLUX Non-Commercial License.
  • API Access shows whether devices away from home can really reach this computer: the relay checks from the internet, through its tunnel and the router's port, every 30 minutes and on Check now, and says which way works, which doesn't, and what to do about it, instead of trusting what the router said.
  • Qwen-Image-Edit-2511 can be set up for editing pictures on computers with 32 GB of memory or more: it changes what's asked, keeps the rest, and edits text in pictures. Once it's set up, edits use it, and new pictures still use your picture model.
  • Speak a message instead of typing it from a phone or tablet: the recording is turned into text with Whisper on your computer, for you to check before sending.
  • A picture sent to a chat with no model that can see it gets an offer to set one up, such as Gemma 3 4B, with its download size. Once it's installed, the question is asked again with the picture, and answered. Before, the answer ignored the picture and pointed to the Models page.
  • Wan 2.1 T2V 1.3B can be set up for clips on computers with less memory than the 5B model needs, at lower quality. It's marked as not yet tested on real hardware.
  • Wan 2.2 T2V A14B can be set up for clips on computers with 32 GB of memory or more: higher-quality clips from a description, more slowly. Clips from a picture still use Wan 2.2 TI2V 5B. It's marked as not yet tested on real hardware.
  • Z-Image-Turbo can be set up for pictures on the Tools page: photorealistic pictures with clear text in them, in a few steps. It doesn't edit, so it's chosen rather than recommended. With it as your picture model, edits use FLUX.2 [klein] 4B when that's set up.

Changed

  • Pictures and clips use stable-diffusion.cpp master-956, which adds support for newer image and video models. A computer already set up updates itself (about 35 MB) the first time it makes a picture or clip, without setting up again.

Fixed

  • Making a picture or clip no longer fails when chat models are loaded: the models in its way are unloaded first, waiting for any answer being written, and a model starting meanwhile waits for the picture or clip to finish. One picture or clip is made at a time.
  • When a picture or clip still fails, Toskar says why in plain words, such as the computer running out of memory or the graphics card stopping, and what to try, instead of the generator's raw error.
  • Speaking a message and reading aloud work when Core is started with a relative --data-dir.
  • When speech can't be set up, speaking a message says so in plain words instead of showing a file path.
  • The Performance page's live figures update every few seconds while a picture or clip is being made, as they do while a model is loaded, instead of once a minute.
  • Asking for "a 10-second video", "a 720p clip", or "a vertical video" makes it that long and that size, as far as the model allows. Before, a clip was always 2 seconds at the model's usual size.

1.8.1

October 10, 2026

Downloads and release notes ↗Docs for 1.8.1 →

Added

  • Word documents (.docx) and PowerPoint decks (.pptx) can be attached to a chat and connected as knowledge: their text is read, a deck slide by slide. Toskar already made Word documents it couldn't read back.
  • A scanned PDF can be attached to a chat: its pages are read with text recognition, which starts as you attach it, so the question finds the text. Before, it could only be connected as knowledge.

Fixed

  • Asking for a file by its name, such as "save it as hours.html" or "write a script saved as convert.py", makes the file, for web pages, JSON, and code as well as documents and spreadsheets.
  • When you decline a command Toskar asked to run, the answer says it wasn't done, instead of handing you the command to run yourself.
  • Gemma 2 9B, Qwen 2.5 14B, and Qwen 2.5 32B now run a profile's topic checks when they're installed, as Gemma 3 4B does: they passed the topic quality set, and smaller models wrongly refused real questions.
  • Toskar recognizes Spanish more reliably, so an answer or a topic refusal is in the person's language instead of being mistaken for Portuguese.
  • A profile kept to one subject takes a second look before refusing a message, so a question about caring for or using what it covers, such as tire pressure for a tire shop, is answered on small models too. The check of the answer reads the question without its "search the web" wording.
  • A recording attached to a message is transcribed before the answer is written, so "what does this say?" gets the words instead of an empty reply. A reply that's only a tool call that can't run is asked for again in plain text, never shown empty.
  • The quality set's stand-in picture is a whole PNG, and its stand-in clip is a WebM, as the real video tool makes.
  • PDFs Toskar makes show text in any of its languages, not only Western European ones: accents, Greek, Cyrillic, and Vietnamese, with Chinese, Japanese, and Korean fonts downloaded the first time a PDF needs one.
  • Every device connected to a computer can reach it away from home through Toskar's relay, not only the one that bought the subscription: devices get the computer's relay token on the home network.

1.8.0

October 10, 2026

Downloads and release notes ↗Docs for 1.8.0 →

Toskar 1.8 opens your AI to more places. Chat portals put an assistant on its own page or on your website, with its own profile, look, limits, and visitors, and topic controls keep an assistant on its subject, checked before and after each answer. Access from anywhere lets your devices reach your computer away from home, through your router or a relay. The Toskar app on your phone gets your computer's notifications and automations, its memories, personalization, and privacy records, and a share sheet that saves pages and files to Knowledge. Chats keep versions when you try again or edit, several chats can be deleted at once, models can be added from any GGUF file, and the computer starts loading a model as soon as a device starts asking. The API stays compatible: the client contract is 1.26, whose new versions only add fields. Binaries and the apt repository are not signed.

Added

  • Apps can delete several chats in one request, POST /api/v1/conversations/delete, with their messages and files. Each person, paired device, and portal guest can delete only their own chats. The web app and the phone app will use it to select and delete chats together.
  • Chat history has Select: tick several chats, or Shift-click a range, and delete them with one confirmation. Deleted chats, one or several, can be brought back with Undo for 10 seconds. Clearing history in Settings deletes in a few requests instead of one per chat.
  • Settings → History can delete chats last used more than 30, 90, or 365 days ago, keeping pinned chats, and says how many before you confirm. Clearing all history is confirmed in the page, which also works in the desktop app.
  • Deleting a chat that an automation posts its results to says so: the automation keeps running, and its results stay in its history.
  • Try again and Edit in every chat, past ones included. Try again can use another model; Edit puts the message back in the box, and ↑ in an empty box edits the last one. Each makes a new version, and ‹ 1 / 2 › switches between them. A retry says when the answer it replaces had made files or run commands, which stay done.
  • Chats keep versions: retrying an answer or editing a sent message adds a version of that point instead of replacing it, with the conversation that followed each version kept. The API answers from any point (retry_of, edit_of, parent_id on POST /api/v1/chat), lists each message's versions, and switches the one shown; the model is sent only the chat up to that point. Existing chats read as one version at every point.
  • Deliberate, first part: a profile can set Deliberate to Always so its answers are drafted three times, independently, on other computers when there are any, and compared by their short final answers. When most drafts agree, that answer is kept. When they disagree, or the answers are too long to compare, each draft is checked against the others and a judge writes the answer, saying where they still disagree. The answer's steps say how it went, and the drafts and checks are kept with the answer for the app to show. It's off by default while it's measured.
  • The web app's profile editor has the Deliberate setting, Compare independent drafts, under Strategy, and Drafter and Judge models under Show all roles. Under an answer, the drafts open to show each draft's model, computer, and final answer, the checks, and the judge.
  • Adding a knowledge source can also add it to profiles (profile_ids), so what you save from a connected device's share sheet is used in chats right away.
  • Add a model from a GGUF file through the API, POST /api/v1/models/import: a file on this computer, copied into Toskar's models folder or used where it is, or the file itself sent from another device. The file's header is checked first, so an incomplete copy or a file that isn't a model is refused, and the model's name, size, quantization, and context length come from the file.
  • Toskar finds models LM Studio, Ollama, llama.cpp, and GPT4All already downloaded on this computer (GET /api/v1/models/import/found), so they can be added in place without downloading them again.
  • Models → Add a model: send a GGUF file from the browser, add one by its location on this computer (copied or used where it is), pick one another app already downloaded, or paste a link.
  • After adding a model, Check it works loads it and asks one short question: Works, with its speed, or why it didn't answer. Benchmarks have the same one-prompt Quick check.
  • A model you added can be renamed and tagged with what it's good at, so Auto sends coding requests to one good at coding. A vision projector found with it (beside the file, or Ollama's) is offered as its image support.
  • A connected device can now see your computer's notifications and automations, run or pause an automation, and open a result in a chat. It can also manage the computer's memories and personalization, and show what left the computer. Making or editing an automation still happens on the computer or in a chat. Seeing what left the computer needs an admin.
  • Chat portals can be for Members only, who chat as themselves after signing in, or for people you invite by name, each with a one-time link that keeps them signed in to the portal. Remove an invited visitor and they stop at once.
  • Put a chat portal on your own website: list the websites that may show it, then paste a chat button script or a frame snippet from the portal's editor.
  • Chat portals have limits: messages per visitor an hour, the longest message, and chats at once, set under Administer → Portals. Portal chats wait for the chats of the people who use Toskar, and come before automations, so a busy portal never slows them down.
  • Read a chat portal's visitors' conversations and see how much it's used under Administer → Portals. Visitors are told the people who run the portal can read their chats, and each portal keeps them for the time you choose, 30 days to start. A portal's chats show as Portal in What left this computer.
  • A chat portal's page at /p/<portal>: only a chat, with the portal's name, logo, colours, theme, welcome, suggested prompts, and footer. A visitor enters with the passcode (or at once when the portal is open), and comes back to their chat on the next visit.
  • Administer → Portals: add a chat portal, choose its profile, tools, memory, language, and who can chat, set its heading, logo, colours, theme, welcome, suggested prompts, and footer with a live preview, copy its link, and turn it off at once.
  • Chat portals, first part: Admins can create portals through the API, each answering with its own profile, tools, memory setting, and language, open on the network or behind a passcode. Visitors who enter become guests who reach only the portal's chat, with their chats private to each of them. The portal page and its settings screen come next.
  • The user guide explains reaching Toskar away from home: turning on access from anywhere, what the reachability line means, the free ways in when the router won't open a port (a forwarded port, Tailscale), and an organization's own relay. A new guide for organizations covers the ways in for hundreds of people.
  • Access from anywhere, first part: an Admin can turn on a listener for paired devices away from home (API Access → Access from anywhere). It serves only what a phone uses, over HTTPS with the computer's certificate, to paired devices' keys, and turns away an address that keeps sending wrong keys. For now it's reached through a port forwarded by hand or Tailscale; automatic setup comes next.
  • Access from anywhere opens its port on the router by itself (PCP, NAT-PMP, or UPnP), keeps it open while it's on, and closes it when it's turned off or Toskar quits. API Access says whether the computer is reachable directly, over IPv6, or at a forwarded address, or why not, such as an internet provider that shares one address among many homes.
  • The Toskar app can hand this computer the relay token from a subscription, and an Admin's device can turn access from anywhere on at the same time, so subscribing on the device is all the setup it takes.
  • Access from anywhere can go through a relay when nothing more direct reaches the computer. An organization that runs its own Toskar relay enters its name and enrollment secret under API Access → Your organization's relay; the computer then enrolls, keeps its sealed address current there, and holds a tunnel open, so paired devices reach it from anywhere with no open port. Traffic stays encrypted from the device to the computer. API Access shows whether the relay is connected, and paired devices learn which relay to find the computer through.
  • Pairing a device now gives it this computer's route secret, for finding it away from home without Toskar's service learning its addresses; devices paired before get it the next time they connect at home. The computer can seal its addresses in a record only paired devices can open, signed with its certificate.
  • The screenshot pipeline makes Microsoft Store screenshots: a microsoft-store form at 1920×1080, captured at 1440×810 so the layout matches the Mac App Store set. The Toskar Pro for Windows listing in toskar-apps uses them.
  • Off-topic attempts, under a profile's topic controls: what Enforce held in the last 30 days, by day and by where it came from, with the messages. Mark as on topic adds one to the example questions. They're kept and deleted with run records.
  • Topic controls on profiles: say what an assistant stays on, what it never discusses, and the reply to anything else. The rules come first in every chat, as the administrator's, and the person's preferences, memories, pasted text, and API instructions can't change them. A chat portal's profile brings them along.
  • Enforce for topic controls: each message is checked before it's answered, and an off-topic one gets the set reply without running the full answer. Each answer is checked too, and one that went off topic anyway is replaced. The profile editor's Topics tab chooses Guide or Enforce, and the run trace records what the check found.
  • Pin an API key to a profile: every request with the key, from a website or another app, answers with that profile and its topic controls, and can't name another profile or a model.
  • Pin Members, Visitors, or one person to profiles from People: they chat only with those profiles, such as an assistant with topic controls, and the chat's profile picker shows only them. The Owner and Admins are never pinned.
  • Quality cases for topic controls: a tire shop with Enforce, its real questions and small talk, and jailbreaks in many forms and languages. The report gives the share held and the share wrongly refused, and the weekly real-model run includes it.
  • Try it, under a profile's topic controls: run a message past them as they are in the editor, or one of a few made-up attempts to talk the assistant off topic, and see the label and the reply before saving.
  • A web limit for topic controls: web search and opening pages keep to the profile's sites, such as a shop's own site, and its words are added to every search.
  • POST /api/v1/models/warm starts loading the model a chat would use and answers at once, so a phone, tablet, watch, or TV can have the computer load it while someone types or speaks, instead of after the question arrives (#498, contract 1.26). It never loads beside a model that's answering, nor one that wouldn't fit beside those already loaded.

Changed

  • The phone app's text says "this device" instead of "this phone" where it means the device it runs on (installing models, storage, errors, connecting, and onboarding), in all ten languages, since the same app runs on tablets and Macs.
  • The README and the user guide name the two apps: Toskar Pro for your computer, and Toskar for phones, tablets, watches, and TVs (yeixio/toskar-apps#123).
  • The apps for phones, tablets, TVs, and watches call the computer app Toskar Pro where they mean it: connecting, finding computers, its version and encryption in Settings, and the setup steps, which no longer name operating systems. In all ten languages (yeixio/toskar-apps#123).
  • The guide for organizations now covers paired devices through a reverse proxy or Cloudflare Tunnel: pair them by the server's name on the office network, and they keep working away, with the proxy's certificate.
  • Enforce's topic checks run on the smallest installed model verified to hold them (Gemma 3 4B so far) when it fits beside the answering model, instead of always on the answering model. Diagnostics and the Topics tab say when no verified model is installed.
  • Topic checks name the subjects a profile never discusses in the off-topic label, and the topic rules say that how someone asks ("search the web for…") doesn't change the subject.
  • The watch app's text for a question the phone couldn't finish while asleep says the answer appears on the watch once Toskar is open on the phone, and a new line says when the phone is still answering, in all ten languages (yeixio/toskar-apps#149).

Fixed

  • A finished answer could briefly show the chat as it was before it, when the reply stream ended before the answer was saved.
  • Chat no longer looks up Toskar's own guide for questions that aren't about it, such as a word problem that happens to mention minutes and notices, or one that says "until you have". The guide is consulted when most of a question's words appear in one of its passages, or when the question names Toskar or asks what it can do.
  • A profile kept to one subject refuses fewer real questions: asking about caring for or using what it covers counts, how someone asks ("search the web: …") no longer sways the check, and a refusal is always in the person's language.
  • Your messages in a chat no longer have an empty band under the text. Edit sits under the message instead: on hover with a mouse, always on touch.
  • Deleting a model only deletes files in Toskar's models folder.
  • /v1/chat/completions honors temperature and max_tokens (and max_completion_tokens) for the answer, on this computer, a paired one, or an external server. Temperature 0 gives the same most likely answer every time. Before, both were accepted and ignored.
  • A connected device says to update Toskar Pro, instead of "not reachable", when the computer's Toskar Pro is too old to save what you shared to its Knowledge.
  • A portal visitor whose earlier chat was gone, such as after retention, starts a new one instead of sending messages that weren't kept.
  • The release screenshots of the Automations form are taken again. The demo server behind them answered "Fill in the details" with an empty object, so the form failed and the 1.7.0 release's screenshots job stopped there.
  • An assistant with topic controls no longer answers questions about Toskar itself, "remember …" commands, or install offers in place of its own rules.

Security

  • Deleting a chat removes its files only when the chat is yours. Before, a request to delete someone else's chat failed but still removed that chat's files.
  • Toskar's pages can no longer be shown in another website's frame, except a portal's page in the websites its Admin lists.

1.7.0

October 8, 2026

Downloads and release notes ↗Docs for 1.7.0 →

Toskar 1.7 is for more than one person and more than one device. People sign in with their own accounts, by password, invite link, an OpenID Connect provider, or a reverse proxy, and roles decide what each can do; chats, memories, files, and automations stay private to their person, and the existing user becomes the Owner. Automations run on schedules, file and folder changes, web pages, feeds, and webhooks, can chain into each other, and can be set up by asking in a chat. Chat can see pictures and watch videos. Traffic between computers and from phones is encrypted, a phone, tablet, or TV connects with a 6-digit code, and every model shows whether it runs on the GPU. The desktop app's menus and tray call it Toskar Pro, its new name. The API stays compatible: the client contract is 1.15, whose new versions only add fields, and a paired computer still on 1.6 keeps working over plain HTTP, marked Not encrypted, until it's updated. Binaries and the apt repository are not signed.

Added

  • Chain automations: one can run right after another finishes, given its result, such as a research automation followed by one that drafts a note from it, or only when the first one notifies, such as acting once a price drops. Choose "After another automation" under "Runs", or toskarctl automations create --trigger after --after <id>. Loops are refused when saved, and a chain stops after five in a row.
  • Continue in chat from any automation run: the result opens in a chat named after the automation, or in the chat it was already posted to, so you can ask a follow-up about it.
  • A daily automation digest: turn it on in Settings → Notifications and pick a time, and once a day one message in an "Automation digest" chat sums up what every automation found and what failed, with one notification instead of many. A quiet day sends nothing.
  • Automations can run when files in a folder in your home folder, or one file, change. The task is told which files were added, changed, and removed, with the start of each new or changed text file. Choose "When a file or folder changes" under "Runs", or `toskarctl automations create --trigger folder --trigger-path ~/Documents/Invoices`.
  • Ask in chat for something on a schedule, such as "every morning at 8, summarize the news", and the answer shows the automation as a card: what it does, when, and when it notifies. Nothing is scheduled until you press Create, and the automation remembers the chat it came from. The request is read in the language it's written in.
  • Profiles have an Automations capability, on in the built-in profiles, for setting up automations from chat.
  • POST /api/v1/automations/parse reads a request such as "every morning at 8, tell me if the price is below $500" into an automation's name, task, schedule, and notification, in any of the App's languages or English. It uses the same words as the Automations page, now kept in i18n/requests/, so the page, toskarctl, and chat can read requests the same way.
  • toskarctl automations parse "<request>" prints what Toskar understood from a request, and toskarctl automations create --request "<request>" creates the automation from it.
  • An automation made from a chat posts each result it notifies about to that chat, so you can reply to it there. The notification opens the chat, and the result is marked with the automation it came from.
  • Automations can also save each result as a Markdown file in a folder you choose in your home folder, such as ~/Documents/Toskar, where other apps and backups find it. The run history shows where each result went. toskarctl automations create|update --save-folder sets it.
  • Automation templates: morning briefing, price watch, back in stock, release notes, page watch, and a new folder summary. Choosing one asks only for what it needs, such as a product link and a price, and fills in the form without reading a request; "Describe it in your own words instead" starts from an example request.
  • Automations can run when a web page changes or a feed has new posts, instead of every time. The schedule says how often to check; a check is a quick fetch with no model, and the task runs only when something changed, told what it was: the lines that came and went, or the new posts. Choose it under "Runs" in the form, or with `toskarctl automations create --trigger page|feed --trigger-url <url>`.
  • Automations can run when another service calls their webhook link: pick "When another service calls its link (webhook)" under "Runs", then make the link on the automation's page. It's shown once, only a hash of it is kept, and making a new one stops the old one. The request's body, up to 64 KB, is given to the task as data. toskarctl automations hook <id> makes a link from the command line.
  • A "Only when started" schedule, for an automation that runs only from Run now or its webhook.
  • Chat can see pictures. Attach a photo, screenshot, or diagram and ask about it: Toskar shows it to a vision model on this computer. If the chat's model reads text only, the largest vision model that fits answers instead, and a follow-up question still sees the picture. With no vision model installed, the answer says how to get one.
  • Two vision models in the catalog: Gemma 3 4B and Qwen 2.5 VL 7B. Each downloads with its image projector.
  • Chat can watch videos. Attach an MP4, MOV, or WebM clip and ask about it: a vision model is shown six frames sampled through it, with the time of each, and can transcribe what's said. Frames are read on this computer with the speech tools, which install the first time they're needed. A follow-up question about the same clip doesn't read it again.
  • A phone can connect with a 6-digit code instead of a copied API key: the computer shows the code (POST /api/v1/devices/pairing), and the phone sends it from the local network (POST /api/v1/devices/pair) to get a key of its own, named after it. A phone's key reaches only chat, its conversations, and the lists the phone reads.
  • Connect a device, on Computers and in API Access: it shows a 6-digit code to type on a phone, tablet, or TV, with the steps and this computer's address, turns on local network access first if it's off, and says when the device has connected. Connected devices are listed under API Access → Devices, where each can be disconnected.
  • Settings → Your data stays private says whether this computer's disk is encrypted: "Your disk is encrypted with FileVault", or, when it isn't, where to turn on FileVault, BitLocker or Device Encryption, or LUKS. It checks the disk that holds Toskar's data, without administrator rights.
  • Running models report where they run, read from llama.cpp's own log when it loads them: on the GPU, partly, or on the CPU, with the device, the layers on the GPU, the GPU memory, and the reason when a model isn't fully on the GPU. GET /api/v1/models/running has it in acceleration, and GET /api/v1/health sums it up in acceleration without changing status (#317).
  • Performance › Activity marks each reply GPU or CPU, with the device in its details, and filters by it; benchmark results say where each model ran. The user guide has a section, "Is my model using the GPU?", which the assistant also answers from (#317).
  • The Linux packages recommend the Vulkan loader, Mesa's Vulkan drivers, vulkan-tools, and pciutils, so apt and dnf install what an AMD or Intel GPU needs, and the service user joins the render and video groups so it can use the card. install.sh prints the command to install NVIDIA's driver when it finds an NVIDIA card without it, and install.ps1 points to the card maker's driver when Windows has no Vulkan (#317).
  • The Dockerfile has a gpu target with Vulkan and Mesa's drivers, for a container given the card with --device /dev/dri.
  • docs/gpu.md covers what each computer needs for its GPU, what the installers set up, and how to fix a model that runs on the CPU.
  • GET /api/v1/performance/live: each computer's CPU, memory, and GPU figures (busy %, video memory, temperature, power) now, over the last hour, and as per-minute averages over the last day, on Linux, macOS, and Windows without root or admin rights. A figure a computer can't give is left out rather than shown as 0. Paired computers' figures are included (#317).
  • Diagnostics lists what this computer still needs for Toskar to use its graphics card, each with the command that fixes it here: the Vulkan loader or drivers (for this Linux distribution's package manager), permission for Toskar's user to open the card, NVIDIA's driver, a Windows card's driver, or the CPU-only llama.cpp. GET /api/v1/diagnostics/gpu returns the same list (#317).
  • A status on each running model says where it runs: green on the GPU, amber when only part of it fits, red when the computer has a GPU it isn't using, and grey on a computer without one. Tap it for the device, the layers on the GPU, and the GPU memory, and for anything short of green, why and what to do. It's on Models › Running and on the Performance overview (#317).
  • The Performance overview shows each computer's live CPU and memory, and each GPU's use, video memory, temperature, and power where the computer gives them, with the last hour as a line. Diagnostics has a GPU acceleration row.
  • Replies, run traces, and benchmark samples record the backend and device that produced them, such as Vulkan on an AMD Radeon RX 7900 XTX, so performance history can tell GPU replies from CPU ones. GET /api/v1/runtimes lists what the installed llama.cpp build can run on (#317).
  • Vision models installed from Hugging Face can see pictures. Browse all models marks a repository that ships an image projector as a vision model. Installing it downloads the projector with it, and the model's download is checked against the SHA-256 the repository lists. Reinstalling a vision model installed from Hugging Face before now adds its projector.
  • Image and video generation say up front what a computer can do. A computer without the memory for a model isn't offered it, can't set it up, and is told how much it needs; one that's short of what the model is comfortable with is warned it may be slow or fail. Without GPU acceleration (every build but macOS), the setup offer, the Tools page, and the progress line say a picture takes a few minutes and a clip can take most of an hour.
  • A saved answer's metadata records the backend and device that ran it, when it ran on that computer, so the iPhone app can show GPU or CPU for answers from a paired computer too. The client contract is 1.8 (#317).
  • Everyone can connect their own phone, tablet, or TV from Settings → Your devices, and see and disconnect the devices connected as them. A device chats as the person who connected it. Only Admins and the Owner turn on local network access to do it.
  • The start of people and roles: this Toskar's one user becomes its Owner, and API keys, chats, memories, files, and automations record whose they are. GET /api/v1/me says who a request is from. Nothing changes for one person; signing in other people comes next.
  • Notifications belong to their person. Members see their own automation results and approval requests in the bell, and nobody else does; Admins and the Owner also see notices about models, training, and health. Email, push, webhooks, and desktop notices still go only to the Owner's destinations, so they never carry another person's results.
  • Sign in with an OpenID Connect provider such as Google, Microsoft Entra ID, Okta, Authentik, Keycloak, or Authelia. Set TOSKAR_OIDC_ISSUER, TOSKAR_OIDC_CLIENT_ID, and TOSKAR_OIDC_CLIENT_SECRET, and the sign-in screen offers it. People are made the first time they sign in, with a role from their groups.
  • Each person chooses their own App language, the language answers are written in, and Personalization, starting from the Owner's. Members and Visitors find them in Settings, and their chats and automations follow their own choices.
  • Chats, memories, files, and automations are private to their person, ready for more than one person to use Toskar: nobody else lists, reads, or changes them, a chat draws only on its person's memories, and an automation runs as its person. Nothing changes while you're the only one.
  • Sign in through a reverse proxy such as Authelia, Authentik, Cloudflare Access, or Tailscale. List the proxy in trusted_proxies (TOSKAR_TRUSTED_PROXIES), and Toskar takes the person from its Remote-User header, makes them the first time, and gives them a role from their groups. The headers count only from the listed addresses.
  • Roles now decide what each person can do. Visitors chat; Members also have their own memories and automations; Admins run everything else, such as models, tools, knowledge, computers, keys, people, and settings; and only the Owner can erase everything. The app shows each person only the pages and controls their role can use, and the service refuses the rest.
  • People can sign in. Admins and the Owner add someone and get a one-time link; opening it, they choose a username and password and are signed in. Passwords are hashed with argon2id, and too many wrong ones make that username wait. Admins add Members and Visitors, the Owner adds Admins too, and disabling someone signs them out everywhere. A phone paired with a code now gets the key of whoever showed the code. The app's sign-in screen and People page come next.
  • The app has a People page under Administer. Admins and the Owner add someone, pick a role, and copy a one-time link to send them; the page also changes roles, makes new password links, and disables people. Opening the link, the person chooses a username and password and is signed in. From another device, the app now asks for a username and password, or an API key, and Settings shows who is signed in with a Sign out button.
  • Connect a device shows the code of this computer's certificate. A phone on the new app connects over HTTPS, saves that certificate, and shows the same code in its Settings, so you can see it reached this computer.
  • The automation form sets up the new schedules: pick days of the week (such as Monday to Friday), add more times a day, run every month on a day of the month, or enter a cron expression. Schedules read as people say them, such as "Weekdays at 9:00 AM" or "Every Monday and Friday at 8:00 AM and 5:00 PM", in every App language.
  • Automations can run on several weekdays, such as weekdays only, at several times a day, monthly on a day of the month (a shorter month uses its last day), or on a cron expression. The API takes them as weekdays, times, month_day, and cron, and so does `toskarctl automations create --schedule monthly --day 1 or --weekday weekdays --at 08:00,17:00 or --schedule cron --cron "0 9 * * 1-5"`. The model that reads requests the words can't can now give these schedules too.
  • Every page in the sidebar has a screenshot, and the README shows them. Train, Knowledge, Memory, Tools, and Profiles & Orchestration now have demo data in the screenshot server (a training run in progress, connected documents, memories, a Notion tool source with connected services, and a team profile across two computers), so the release attaches screenshot-train.png through screenshot-profiles.png next to the others (#304). The README stills and walkthrough show the new sidebar.
  • Image and video generation use the GPU on Linux and Windows. With a graphics card that has a Vulkan driver (NVIDIA, AMD, or Intel), setup installs stable-diffusion.cpp's Vulkan build, so a picture takes seconds instead of minutes. A Vulkan build that can't start on the GPU falls back to the CPU build by itself. The Tools page says which build makes them, with a button to switch. An existing setup on the CPU build offers **Use the GPU build** there.
  • Paired computers that make pictures on the GPU are now reported as accelerated, so heavy image work goes to them.
  • A chat can offer to set up image or video generation on a paired computer. When a laptop would make pictures slowly on its CPU and a paired workstation has a GPU for it, "Make me a picture of…" offers the workstation, with its free disk space. Set up starts the download there and shows its progress, and once it's ready the picture is made there. A computer without the memory or the disk space isn't offered.
  • site/highlights.json holds the latest minor release's highlights, which toskar.ai shows as "New in 1.6" instead of a list written into the site. A new minor release updates it in the same pull request: scripts/changelog.py check fails until its version matches the latest release in CHANGELOG.md.
  • Toskar says when a newer version is out. Release builds installed from a download look at toskar.ai once a day, sending nothing but the request, and Settings › About shows the new version with what's new and a download link. Turn it off in About; each check is listed in What left this computer. The App Store edition and the desktop app's copy update with their app, so they don't check.

Changed

  • API Access, Performance, and Diagnostics match the rest of the app: the app picker under Use Toskar in other AI apps and Performance's views are segmented tabs, buttons share the same sizes, and Revoke on an API key is red, as a destructive action. Diagnostics is titled Diagnostics, as in the sidebar.
  • The Automations page reads a request on the computer, so the page, toskarctl, and chat read requests the same way. When its words can't find a schedule, such as "first thing on weekdays", a model reads the request, and the form says so, so you can check it before saving.
  • The computer adds what an automation's condition needs, such as the price in the threshold's currency, when each run starts, so automations made with toskarctl, chat, or the API check their conditions the same way as ones made on the page. A saved prompt is only the task.
  • A run's history says why it did or didn't notify as the computer decided it, instead of the page guessing from the result. Runs record it as notify_detail and notify_values.
  • Automations explains itself. With none yet, the page shows the three parts of an automation (when it runs, what Toskar does, and when you hear about it) and five ideas to start from: a morning news brief, a price watch, a back-in-stock check, weekly release notes, and a page watch. Choosing an idea opens the form already filled in. With some automations, How it works shows the same.
  • The new-automation form has two numbered steps, Describe it and Check the details, and the button that reads your description is now Fill in the details. A Summary above Create automation says when it runs, what it does, and when you hear about it. Test run says that it runs once without saving.
  • The Automations list uses the same filters, search box, and status chips as the rest of the app.
  • Chat has a cleaner look. Messages and the message box share one centred column; your messages sit in a soft bubble and replies read as plain text beside the Toskar mark. The header keeps History, the chat's name, Delete, and New chat in one row, and the message box starts small and grows as you type.
  • The chat history panel matches the new Chat look: icon buttons to dock and close it, a search box with a search icon, one line per chat, and tidier menus.
  • The context gauge beside Send is an open meter that fills as the chat uses the model's memory and turns amber near the limit, so it no longer looks like an empty radio button.
  • The Computers page is titled Computers, as in the sidebar. While this is the only computer, it explains how to add another in three steps: install Toskar there, connect them (Available to add on the same network, or Add by command for a server), and share the work. How to add a computer shows the same later. Computer cards use quieter Manage buttons and a menu like the rest of the app.
  • The desktop app can post desktop notices itself: with TOSKAR_DESKTOP_NOTIFICATIONS=shell, which the app sets on the service it starts, notices go to it as notification.desktop events, so they carry Toskar's name and icon, open the app when clicked, and work in the Mac App Store edition. Without it the service posts them as before.
  • The same card now says that traffic between your computers, and from other devices on your network, is encrypted.
  • Knowledge explains itself and shows where each source is used. Every source says which profiles' chats use it; one that no profile uses says that Chat doesn't look in it, with Use with to add it to the profile chats start with. Connecting a source asks which profiles should use it, with that profile already checked, so it works in Chat straight away. With nothing connected yet, the page explains knowledge in three steps and shows what people connect; How it works shows the same later.
  • Turning local network access on or off applies at once, without quitting and reopening Toskar. Turning it off also disconnects other devices right away.
  • Turning on local network access no longer asks apps on this computer for an API key. Requests over loopback (127.0.0.1 or ::1) work as before, so the desktop app, toskarctl, and a browser on this computer keep working. Other devices still need a key. A request carrying proxy forwarding headers (Forwarded, X-Forwarded-For, X-Real-IP) counts as coming from the network.
  • Memory explains itself. With nothing remembered yet, the page shows how memory works in three steps (tell it once, it's used when it fits, you stay in control), says that documents and data belong in Knowledge, and offers example memories that fill in the Add box. How it works shows the same later. Use memory in chats is now a switch.
  • Each memory shows its words, where it came from, and Paused or This computer only when they apply. Edit changes the words and the category; Pause, This computer only, and Delete are in the memory's menu.
  • Models has a cleaner look. Discover, Installed, and Running are one segmented control beside a search box with a search icon. Each model's fit shows as a coloured chip (green when it runs well, amber when it's tight, red when it won't), and every action in a card is the same size. The "Models for" strip is tidier, and on Running, Stop sits beside Open chat with a bar for how much of the computer's memory the model holds.
  • Local network access no longer needs an API key once someone besides the Owner has been added, since they sign in with a password.
  • The profile editor opens across the whole page and is split into four tabs: Models, Tools, Knowledge and memory, and Strategy and computers. It was one long column inside a half-width card. Models shows the roles in use, with Show all roles for the rest; Tools uses switches, with per-tool permissions folded below them.
  • The real-model quality run rides out a daemon that goes away for a moment (a restart or a settings change), retrying for up to three minutes. If the daemon starts asking for an API key partway through, the run stops with one clear reason instead of failing every case left. A case that stops before it is checked now counts as failed in the summary.
  • The Docker image takes a COMMIT build argument, which /api/v1/version reports.
  • The README's Quick start shows how to use Toskar from other devices on your network: turning on local network access and creating a key, restarting, finding the address, and the settings for a server without a browser.
  • CI can run on self-hosted runners: the repository variables CI_RUNS_ON (pull request and push checks) and TRUSTED_RUNS_ON (releases, screenshots, and the small-model quality job) pick the runner, and GitHub's runners stay the default. The installer test always runs on a GitHub runner, since it installs system-wide.
  • Settings keeps your own preferences, and the rest moved to where it's used (#203). Find other computers and the External server (advanced mode) are on the Computers page, under Network & access. Connected services (GitHub, Home Assistant, email, and calendar) are on the Tools page, next to tool sources. The API access summary card is gone from Settings; API Access already manages it. Settings' Network & access group is now Notifications.
  • The sidebar is grouped by who uses each page (#203). Use has Chat, Automations, Knowledge, and Memory; Customize has Models, Tools, and Profiles & Orchestration; Administer has Train, Computers, API Access, Performance, and Diagnostics; Settings stays at the bottom. Tools and Profiles are always listed: advanced mode no longer hides whole pages, only expert controls inside them. Administer is closed until you open it, and stays as you left it; it opens by itself while one of its pages is showing. Page addresses are unchanged.
  • The iPhone app's quality run now fails when its copies of core's text rules (what needs the web, small talk, search queries, memory commands) or its quality cases fall behind core's.
  • The Tools page has three tabs: Your tools, Add more, and Images and video. Your tools lists every tool in a compact grid with search and filters; Add more has connected services and tool sources; Images and video sets up image and clip generation. How it works explains that tools are built in or added, and that each profile chooses which tools its chats may use, with a link to Profiles & Orchestration. A link to /tools?tab=add opens Add more.
  • The Train page is titled Train, as in the sidebar, and its buttons match the rest of the app. Try an example and Build a new AI appear once: in the introduction when no AI is open, and at the top when one is.
  • Files that belong to no chat are removed after a week, instead of piling up forever. That covers files an automation run, the API, or an MCP client made, and uploads never sent with a message. Files in a chat stay with the chat. The API shows when such a file goes as expires_at.
  • TOSKAR_WEB_FIXTURES, for quality runs only, makes the daemon answer web search, places, and page reads from a file of fixed pages (tests/quality/web.json) instead of the internet, and log a warning that it does. The self-hosted real-model run uses it, so its answers are checked against the same pages as every other run.
  • Chat looks things up on the web before answering when a message asks it to find something: a recommendation, where to buy, what something costs, reviews, a result or a release, or "check" and "find" for me. Before, these were answered from the model's memory, which for a small model often meant naming products that don't fit or telling you to search yourself.
  • A follow-up such as "Can you provide a link to that bike?" is searched by what it refers to: the model writes the search from the conversation, so it names the bike.
  • An answer that still tells you to search, check websites, or pretends to browse is looked up and written again from what the web says. Small talk such as "Hi! How are you?" is never looked up.

Removed

  • LLaVA 1.6 Mistral 7B left the catalog. It downloaded without its projector, so it couldn't see pictures in Toskar.

Fixed

  • The settings to paste into another AI app can be scrolled sideways from the keyboard.
  • "Notify on change" notifies when something actually changed, not when the model words the same facts differently: it compares the values the automation tracks and the pages it read, and otherwise asks the model whether anything meaningful changed. The notification says what changed.
  • "Notify when available" works in every language: the result now carries an availability flag instead of being read for English phrases.
  • The Folder summary template now watches its folder, so its runs see the files; before, a run couldn't open files outside Toskar's workspace.
  • An automation's history loads its newest 20 runs, with Show older runs for the rest, instead of every run it ever made.
  • Preview answers in the automation's chosen language, and with its schedule's time zone, as the saved automation will.
  • Automations run two at a time, so a slow one no longer holds up the others, and a run that takes more than 20 minutes is stopped instead of hanging.
  • Run now starts the run and returns right away; closing the page no longer stops it, and its result appears when it finishes.
  • An automation that fails three times in a row, for any reason, is paused with a notification saying why, in the App language. Retries go by the kind of error rather than its English wording.
  • On Linux, the hardware check counts a memory limit set on Toskar's container or service (cgroup v2 memory.max, such as docker run --memory or systemd MemoryMax=), so model recommendations fit the memory Toskar may use instead of the whole machine's.
  • An open chat has a page heading for screen readers.
  • In the dark theme, dialogs and drawers dim the page behind them instead of covering it in a grey haze.
  • Chats no longer pile up as active tasks. A chat's task was never moved past pending, so the Performance overview counted every chat ever sent as an active task and listed the oldest as Pending. A chat's task now ends completed, failed, or cancelled with the reply, and tasks left unfinished from before are settled when Toskar starts. Chats still raise no "task finished" notice.
  • Remove from team asks first; it removed the computer at once.
  • The Remove button for a model under Manage is readable in the dark theme.
  • The copper Ratatoskr label in dark mode is a little lighter, so it meets the 4.5:1 contrast small text needs. The mascot's colors are unchanged.
  • Profiles you made before image generation existed couldn't make pictures: a tool a profile doesn't list counts as off, and only the built-in profiles gained new tools. Your own profiles now gain new tools that stay on this computer and only read or make files in Toskar's store (pictures, clips, audio, files, drafting automations). Web, browser, terminal, file writes, and Git still wait for you to turn them on, and a tool you turned off stays off.
  • Asking for a picture in a chat whose profile keeps image generation off says so, and where to turn it on, instead of listing websites. A request for a picture or clip no longer searches the web first.
  • A chat's messages always load in the order they were sent. A reply could appear above the question it answered when both were saved within the same instant.
  • On Windows, a graphics card's memory is read from the driver's registry entry, so cards with more than 4 GB are no longer reported as having 4 GB or less, and model recommendations use the real figure. Only graphics cards that are present are listed, so a card that was removed from the computer does not keep being reported.
  • On Windows, a build from source now shows the web UI. The page used to stay blank because the scripts and styles under /assets/ were answered with the start page instead of the files.
  • On Windows, make start from PowerShell or cmd now builds bin/toskar.exe and runs it, and no longer prints "The system cannot find the path specified." when it looks up the git commit.
  • Gemma models answer the question instead of the instructions. Gemma has no place for system instructions, so Toskar's (the date, the reply language, how to answer) reached it unmarked, as if the user had written them. Gemma 3 4B answered "Say hi in three words" with "You've provided the current date and time…". The instructions are now marked as coming from the app.
  • A running model's speed is filled in from its latest replies, and its device is the one it runs on rather than the first graphics card found.
  • Browse all models offers each repository's real model file instead of a guessed name. It skips speculative-decoding draft files, split parts, and extras in subfolders.
  • Asking chat for a picture or a clip makes it. "Draw a dog", "please generate an image of a dog", or "are you able to make a picture of a dog for me?" used to get "I can't draw" or a list of stock-photo sites from a small model, though image generation was set up. Toskar now makes the picture itself, with the model only writing what to draw, and the same for changing an attached picture or making a short video.
  • Pictures, clips, audio read aloud, files code saved, and page screenshots are attached to the answer that made them; only files created as documents were before.
  • A request for a picture asked as a question, before image generation is set up, gets the setup offer with the request, so it's finished once the model is ready, instead of only "No, I can't generate images right now".
  • On Linux, AMD and Intel graphics cards are found from the kernel's card list, without lspci, and an AMD card's memory is counted, so model recommendations use it. A card counts only when its render device is present, so a container without the card isn't told it has one.
  • On Linux and Windows, llama.cpp now installs its Vulkan build when the computer has a GPU with a Vulkan driver, instead of always the CPU-only build. Chats that took minutes on the CPU now run on the graphics card. Computers without a usable GPU still get the CPU build.
  • A llama.cpp install that has the CPU build is swapped for the GPU build the next time Toskar starts, or when you install llama.cpp again from the Runtimes page.
  • llama.cpp reports a GPU backend only when the installed build has one, so Toskar no longer says it can use the GPU when it cannot.
  • Installing llama.cpp again now replaces the whole install. Before, a reinstall could leave the old build in place.
  • Running models say whether they're on the GPU again with current llama.cpp builds, which log the devices and offloaded layers only at trace verbosity and in a new format. Toskar now asks llama-server for that verbosity and reads the new format (#317).
  • A model that can't call tools, such as Gemma 2, gets current answers again. Choosing one turned off every tool, including the web, places, and service look-ups Toskar runs itself before the model answers, so it answered from memory ("I don't have access to real-time information"). Those look-ups now run for every model; a model that can't call tools just isn't asked to.
  • Read aloud in the Mac App Store app reads answers with your Mac's own voices, instead of trying to install speech and showing a sandbox error. The daemon now recognizes the App Sandbox from its folder too, so speech, text recognition, the code environment, image generation, training and the browser no longer try to download what the sandbox can't run.
  • When speech can't be set up, Read aloud says so in plain words, without file paths.
  • go test ./internal/hardware/ passes on Windows: the Linux GPU tests, whose fake device paths Windows can't create, are skipped there (found by @black-operative).
  • Links inside sentences show their text again: the Knowledge page link under a profile's knowledge, the install links for Node.js, uv, and Docker when adding a tool source, the Models link when a specialized AI's base model is not installed, and Open Chat after deploying one.
  • "My daughter's birthday is May 3, remember that." is saved as a memory. Asking at the end of a message used to reach the model, which promised to remember without saving anything.
  • Model names on the Installed and Running tabs are second-level headings, so screen readers no longer skip a level.
  • A file listed under an answer's Sources, attached or made in the chat, can be downloaded from its chip, like the files under Files.
  • "Make a PDF of that" and "can I have it as a Word document?" make the file, instead of telling you to paste the text into a word processor. If an answer still tells you to make a file yourself, Toskar asks the model once more to make it.
  • Only the person who showed a device code sees whether a device connected with it, or cancels it.
  • The live event stream no longer tells one person about another's chats, tool calls, memories, automations, or notifications: each person hears only their own, and everyone still hears about the computers and models.
  • A long chat's summary is made as the person whose chat it is, so it works for people besides the Owner.
  • "Forget that …" found nothing when the exact text didn't match, because its fallback search read the wrong columns.
  • A request for a picture with the word misspelled, such as "make a picutre of a dog", makes the picture. When a request is worded in a way Toskar doesn't recognize and the model says it can't make pictures, or points to DALL-E, Midjourney, or ASCII art, Toskar makes the picture anyway.
  • "Make" in a message no longer offers the terminal unless it means the build tool ("make test", a Makefile), so "make a picture of a dog" can't turn into a command that draws ASCII art.
  • Saving a profile without changing its tools no longer turns tools on. The editor showed every tool a profile didn't list as Always allow, though the daemon treats them as off, so saving gave the profile every tool, including the terminal and git push. It also kept tools from connected services that the editor doesn't list, which saving used to drop.
  • A question asking for a figure, such as "How much does it cost?", reads the best page it finds even at Fast effort, instead of answering from search snippets that rarely carry the price.
  • With a tool that changes things available, the assistant calls it (it asks you first) instead of saying it can't and telling you to run a command such as rm -rf yourself.
  • The Docker image includes the libraries the llama.cpp runtime needs (libgomp1, libcurl4), so a model the container installs can start.
  • The source-type badge on the Knowledge page ("Linked", "Database") now meets WCAG AA contrast in the light theme; it measured 4.43:1.
  • In the dark theme, red labels and buttons are easier to read: the red is a little brighter, so text such as a blocking problem with a training example meets the contrast standard on its tinted background.
  • A training run's progress bar has a name for screen readers, and the example request on the Deploy step can be scrolled from the keyboard.
  • A question about the weather in a month or season, such as "Juneau in June", is answered with typical figures from climate pages, not today's reading. If a word could be a month or part of a place name, the answer says how it read it.
  • What the model writes before using a tool, such as "Let me find that information for you.", no longer appears in the answer.
  • A weather answer links the wttr.in forecast page, such as wttr.in/juneau, instead of the percent-encoded query Toskar reads current conditions from. The Sources list does the same.

Security

  • A run started by what a trigger delivered (a webhook's body, a changed page, new feed posts, or changed files) can't use tools that change things outside Toskar, even ones approved for that automation: someone else wrote what it read. Those tools are skipped and reported, as unapproved tools are.
  • Traffic between your computers is encrypted. Bifrost speaks TLS, and each computer is checked against the key it paired with, so another machine at its address is refused. Chats placed on another computer, their answers, tool jobs and their files, and training runs no longer cross the network readable. Once a computer has used encryption, it is never reached without it again.
  • A paired computer still on an older Toskar keeps working over plain HTTP, and the Computers page marks it Not encrypted until it's updated.
  • The API on port 7331 answers HTTPS. Phones, browsers, and apps on your network can use https:// with the same port, so keys and chats no longer cross the network readable. Toskar makes and keeps its own certificate, and API Access shows the start of its fingerprint to compare with what a browser shows. You can use your own certificate with api_tls_cert and api_tls_key (or TOSKAR_API_TLS_CERT and TOSKAR_API_TLS_KEY). Plain HTTP still works from this computer and, for now, for phones on an older app.
  • Websites open in a browser on this computer can no longer use or read the Toskar API on 127.0.0.1. The daemon now answers browsers only from its own web UI, the desktop app, and loopback addresses on its own port, or when the request carries an API key, and no longer sends Access-Control-Allow-Origin: *. It also refuses addresses that are not this computer's over loopback, which blocks DNS rebinding. toskarctl, the iPhone app, and other devices with a key work as before.

1.6.1

October 4, 2026

Downloads and release notes ↗Docs for 1.6.1 →

This release carries the iPhone app's text for its on-device chat, so the iPhone release built from it can show it. The engine, API, configuration, and data are unchanged from 1.6.0: the client contract stays 1.6, and pairing works with computers on 1.5.0 and 1.6.0. Binaries and the apt repository are not signed.

Added

  • The iPhone app's text for on-device chat: looking up current information on the web with its sources, Auto choosing a model on the phone and falling back to a smaller one, the context ring and its breakdown, memories kept on the phone ("Remember that …"), and their switches in Settings. It is in English; other languages show English until they are translated.

1.6.0

October 3, 2026

Downloads and release notes ↗Docs for 1.6.0 →

Yggdrasil is now Toskar: the engine, programs, packages, repository, and Go module (github.com/yeixio/toskar-core) take the new name, and Yggdrasil stays in the story as the world tree. Chat reopens the conversation you left, with its context gauge, and the iPhone and desktop apps get the text for Auto, chat options, and a clearer message when another copy is running. Existing installs keep working without changes: their data folder and database stay where they are, and the old program names, YGGDRASIL_* variables, API keys, headers, MCP names, and Linux service name still work. The API, configuration, and data are compatible with 1.5: the client contract is 1.6, which adds an optional field, and pairing works with computers on 1.5.0. Go code that imports core must use the new module path. Binaries and the apt repository are not signed.

Added

  • The desktop app's catalog has the text it shows when another Toskar already runs on this computer at the address it needs, in all ten languages, so the app can say so instead of quietly showing the other one's models and chats.
  • The iPhone app's chat text has Auto, the chat options (profile, run on, effort, and memory), and the model Auto used, in all ten languages, so the iPhone can offer the same chat choices as the web UI.

Changed

  • The client contract is 1.6: an answer's metadata can carry context, how full the model's window was for that answer, so clients can show the context gauge when a chat is opened again.
  • Release archives are named toskar-<version>-<os>-<arch>-headless.tar.gz. The install scripts put Toskar in ~/.local/lib/toskar or /usr/local/lib/toskar on macOS, run by the launchd service ai.toskar.toskar, and in %LOCALAPPDATA%\Programs\Toskar on Windows, started by the scheduled task Toskar. Installing over a setup from before the rename removes its old service or task, so only one copy runs, and leaves the old install folder as a link to the new one. The scripts can still install a release from before the rename.
  • The programs are now toskar (the daemon) and toskarctl (the command line). yggdrasil-daemon and yggctl are installed as links to them, so launchd agents, systemd units, scripts, and MCP settings that run the old names keep working; on Windows, the install script adds them. Shell completion works for both names. The join command a computer prints still says yggctl join, so it works on a computer with an older version.
  • The Linux desktop entry and icons are named toskar (toskar.desktop, apps/toskar.png). The logo files in docs/brand are named toskar-*; the mark, Yggdrasil the world tree, is unchanged.
  • The web UI keeps its saved API key, language, effort, and settings in this browser under Toskar names, and moves them from their Yggdrasil names the first time it opens, so nobody is signed out or loses settings. Signing in to a tool source in a separate window works with computers on either side of the rename.
  • The app says Toskar, in all ten languages: the window title, the sidebar, onboarding, settings, notifications, errors, and the iPhone app's text. Yggdrasil stays the world tree in the lore: the stories of the logo, Ratatoskr, the Norns, Mímir, and Odin. The iPhone app's About text says where the name comes from: Ratatoskr, who carries messages along the world tree.
  • New installs keep their data in a Toskar folder (~/Library/Application Support/Toskar, %LOCALAPPDATA%\Toskar, or ~/.local/share/toskar) and database (toskar.db). An existing install keeps using its Yggdrasil folder and yggdrasil.db where they are; nothing is moved.
  • Environment variables have new names starting with TOSKAR_, such as TOSKAR_API_PORT, TOSKAR_API_KEY, and TOSKAR_URL. The YGGDRASIL_ names keep working, so existing Docker, systemd, launchd, and MCP settings need no change. When both are set, the TOSKAR_ name wins. The install scripts accept both names.
  • Messages from the daemon, toskarctl, the install scripts, and notifications say Toskar: errors, the ntfy and desktop notification titles, email subjects, and the version notice (Toskar Core). The assistant answers questions about itself from Toskar's user guide, and still recognizes them when they say Yggdrasil. The product field in /health, /about, and /version stays Yggdrasil within contract 1.x.
  • API responses carry a Toskar-Contract header beside Yggdrasil-Contract, and clients may send Toskar-Client-Contract or Yggdrasil-Client-Contract. Webhooks are signed in Toskar-Signature and Yggdrasil-Signature, and carry Toskar-Notification-Id and Yggdrasil-Notification-Id, with the same values, so existing receivers keep working.
  • Over MCP, Toskar introduces itself as toskar, both as a server to other AI apps and as a client to tool sources, and signs in to services as Toskar. API Access → Use in other AI apps gives settings that name the server toskar; entries made earlier under yggdrasil keep working. A tool source can't be named toskar or yggdrasil.
  • Community ratings fall back to the public summary at its new home, yeixio/toskar-model-data, when the ratings service can't be reached. Earlier versions keep working through GitHub's redirect from the old name.
  • Requests to other services (ratings, place search, Hugging Face, web search, webhooks, ntfy, remote memory, and downloads) identify themselves as Toskar in their User-Agent.
  • The repository is now yeixio/toskar-core, and the Go module path is github.com/yeixio/toskar-core. Old links, clones, release downloads, and the apt source keep working through GitHub's redirect. New Homebrew installs use brew tap yeixio/toskar https://github.com/yeixio/toskar-core; a tap added as yeixio/yggdrasil keeps working.
  • The OpenAI-compatible API takes the assistant's controls in a toskar object. The yggdrasil object still works, and toskar wins when a request sends both. Responses and streamed chunks carry the answer's sources, steps, and progress under both names.
  • The Linux packages are now toskar (deb and rpm), and the service is toskar.service, which also answers to yggdrasil.service. apt-get upgrade and dnf upgrade move an existing yggdrasil install over, keeping its data, its system user, and its service running; a small transitional yggdrasil deb makes that work with apt and can be removed afterwards. The service's log is in journalctl -u toskar.
  • The Homebrew formula is now toskar, and brew upgrade moves an existing yggdrasil install to it.
  • Yggdrasil is now Toskar. The engine is Toskar Core, and the apps are Toskar Desktop and Toskar Mobile; the site is toskar.ai. Yggdrasil stays in the story, as the world tree Ratatoskr runs along, and Toskar is named after him. An existing install keeps working without changes: its data folder and database, YGGDRASIL_* environment variables, yggdrasil-daemon and yggctl, ygg_ API keys and ygj_ join tokens, the Yggdrasil-* headers, the OpenAI yggdrasil object, MCP settings that name yggdrasil, the yggdrasil.service name, and pairing with computers on 1.5.0 all still work. The product field in /health, /about, and /version stays Yggdrasil, and owned_by in /v1/models stays yggdrasil, within contract 1.x.
  • The docs, README, user guide, API description, and project policies say Toskar, and links point to toskar.ai.

Fixed

  • Going to another page and back to Chat reopens the chat you were in, instead of starting a new one, unless you were away for more than 30 minutes. The context gauge keeps its reading when you come back, and after a restart it shows the reading saved with the chat's latest answer.

1.5.0

October 3, 2026

Downloads and release notes ↗Docs for 1.5.0 →

Yggdrasil in ten languages, a UI that meets WCAG 2.2 AA and works on a phone's browser, an assistant that answers from its own user guide and knows the date, a context gauge that shows memory, an optional OpenAI-compatible external server, and fixes for leaks found by new checks. Existing API routes, configuration, and data are compatible with 1.4: the changes add routes, optional fields, and database tables, and migrations run automatically. Pairing is not compatible with earlier versions: it is now signed end to end, so update Yggdrasil on every computer before pairing them or using them together. The removed Orchestrators page redirects to Profiles & Orchestration. Binaries and the apt repository are not signed.

Added

  • Answers with code are checked before they're shown: Go, JSON, Python, and shell code is parsed (never run), and code that doesn't parse goes back to the model with the errors named. At Thorough effort, long answers are also checked for statements that contradict each other or their sources, and rewritten once when they do. The answer's steps say what was checked, and its note says what's left.
  • Describe an automation in the App language, not only in English: German, Spanish, French, Italian, Brazilian Portuguese, Japanese, Korean, and Simplified and Traditional Chinese. Schedules, weekdays, times of day, 24-hour times (18:30, 18 h, 18時, 오후 6시, 下午6点), intervals, one-time runs, and notify conditions all read, and the request box shows its example in the App language. English still works in every language.
  • A price check keeps the currency you wrote, such as 500 €, R$ 2.500, 5万円, or 50만 원, asks for the price in that currency, and shows it in the name and notices. You can change the currency next to the amount. An amount without one is in the App language's usual currency.
  • Run code in a sandbox. code.execute runs Python with numpy, pandas, and matplotlib for calculations, analysis, and charts, with no network, no access to your files beyond those from the chat it is given, and a 90-second limit; charts and files it writes are attached. It runs only in the operating system's sandbox (sandbox-exec on macOS, bubblewrap on Linux) and is unavailable elsewhere; there is no unsandboxed fallback. It asks first by default.
  • Groundwork for Yggdrasil in other languages (#56). The UI's text now comes from a shared translation catalog in i18n/, used by the web UI, the desktop app, and the iPhone app. Settings has an App language: System default, or a language from the catalog; the daemon keeps it (ui_locale), so every app shows the same language. A system in a language without a catalog yet sees English text with its own date and number formats. The navigation is the first part translated. In advanced mode, the en-XA pseudo-locale shows every translated string accented and padded, so text that is not translated or does not fit stands out. Tests check the catalog in CI: valid JSON, no duplicate keys, the same keys, placeholders, and plural forms as English, and no key used in the code missing from English.
  • The desktop app's menus, tray, and closing screen come from the shared catalog (i18n/locales/<language>/desktop.json) and follow the App language; the web UI tells the desktop shell when it changes. With an older desktop app the menus stay in English.
  • The iPhone app's text is in the shared catalog (i18n/locales/<language>/mobile.json), so it can be translated with the rest of Yggdrasil. The app follows the connected computer's App language (ui_locale) unless one is chosen on the phone.
  • Word documents and PDFs. files.create writes .docx and .pdf from Markdown, with headings, styled text, lists, tables, code, and quotes; document.create and pdf.create reach it in that format. PDFs use the standard fonts, so characters outside Western European text show as ?.
  • Spreadsheets with several sheets and formulas: in the CSV for an .xlsx, a line ## Sheet: Name starts another sheet and a cell starting with = is a formula. spreadsheet.create reaches files.create in that format.
  • spreadsheet.analyze summarizes a spreadsheet in the chat column by column (type, counts, minimum, maximum, average, total, or the most common values) with the first rows, so answers can use a whole file.
  • Tool descriptors. Every tool has a version, an input schema, its outputs, a permission level from 1 (low risk, on this computer) to 4 (runs commands or code), its requirements, time limit, provider, and health. GET /api/v1/tools/{id} returns one; the Tools page shows them.
  • Tool audit. Every tool call is recorded with what became of it, how it was allowed, how long it took, and what it was about. GET /api/v1/tools/runs lists them, and the Tools page shows each tool's recent calls. Records expire with run records.
  • Profile strategies. A profile can work Auto (the default), as a Single model, as Planner + workers, or as a Team. Team now runs on the same pipeline as every other chat: a planner splits the request, workers write notes for each part, the answering model writes the answer with tools, memory, and knowledge, and a reviewer checks it. Quick questions are still answered directly. Programming uses the Team strategy.
  • Workers on other computers. With the Team strategy or a worker model, each part of a plan has its own worker, which Norn can place on a paired computer, and parts on different computers are written at the same time. The chat timeline and run details show each worker's model and computer.
  • Model roles. A profile can assign primary, fast, coding, planner, worker, and reviewer models, each with an optional computer. With the chat on Auto, a profile's coding model answers coding requests and its fast model answers quick questions. A fallback order lists the models to try when the answering model fails.
  • Planning: Always asks the planner model to split a request that has no obvious parts.
  • Placement rules per profile. Each paired computer can be Preferred or Never use, and Only this computer keeps every turn here.
  • Profiles & Orchestration. The Profiles page is renamed, and its editor is grouped into Profile, Models, Tools, Memory, Orchestration, and Execution. Each built-in profile has Reset to defaults (POST /api/v1/profiles/{id}/reset), and Duplicate now copies a profile's orchestration and knowledge sources.
  • Run details label each role (Planner, Worker 1, Answer, Reviewer), list them in that order, and count the model calls.
  • Retry on another computer. When a model fails before showing anything, the turn runs again with the same model on another online computer that has it, before trying another model. Retries (1–3) sets how many times.
  • The network advertisement (_localai._tcp) now says where the API is, as api_port in its TXT record. The service's own port is the computer-to-computer port, so an app that finds Yggdrasil on the network, such as the iPhone app, had to assume the default API port.
  • Email and webhook notifications. Settings → Email, push, and webhooks sends notifications to your own SMTP server or to a webhook, with the categories and lowest severity each one receives. Webhooks are signed (HMAC-SHA256, a secret per destination, shown once) and must use HTTPS outside your network; SMTP passwords are kept in the secrets directory. A Test button sends one right away. Deliveries are recorded in What left this computer.
  • Push notifications through ntfy, on ntfy.sh or your own server, for Android, iPhone, and browsers, with no Yeix-hosted service. Severity sets the priority, tapping opens Yggdrasil when its address is set, and on ntfy.sh only a generic notice is sent unless you choose full content.
  • Delivery retries. A failed email, push, or webhook delivery is retried after 1, 5, and 30 minutes, without rerunning the task; a failure that retrying cannot fix stops at once and says what to change.
  • Quiet hours. Desktop notices, email, push, and webhooks wait overnight and go out when quiet hours end; errors still go out unless you choose Hold everything.
  • Health notifications on changes only: a paired computer going offline and coming back, and a model that crashes twice within 30 minutes (at most once an hour, with a hint when it is out of memory). They go to the bell and to destinations that take the Health category.
  • The notification center filters by category, counts repeats ("3 times"), and says when a notification is held or was not delivered everywhere. GET /api/v1/notifications/{id} shows each channel's delivery. The client contract is now 1.1.
  • Browser use: browser.open, browser.extract, and browser.screenshot, and, asking first, browser.click, browser.type, and browser.download, in an isolated headless browser for each chat (the Chrome, Edge, Chromium, or Brave already installed, with a fresh, temporary profile, never your own). Every request is checked so pages cannot reach this computer or the local network, password, payment, and one-time-code fields are never typed into, and each page opened is recorded in What left this computer. A Browser capability turns it on or off in a profile.
  • A shared model rating can include how the model runs on your computer: its typical speed and time to first token, how many of its starts worked, and whether it crashed or ran out of memory. This is a second choice in the rating dialog, off by default, and the dialog shows the numbers before you share. Community ratings show the typical speed and crash rate others reported.
  • With community ratings on, recommendations weigh how people with computers like yours rate each model: a clearly better rated model can be suggested ahead of the usual first choice, and a poorly rated one gives way. Ratings from only a few people count less, models that can't run on your computer are never suggested because of ratings, and a model chosen this way says so.
  • Community model ratings: rate a model with 1 to 5 stars and optional reasons from its card, the Installed tab, or a prompt in chat after you've used it a while. Ratings stay on this computer unless you choose to share one, and the rating dialog shows exactly what sharing sends: the model, your computer's class, and a random rating ID. With Show community ratings on in Settings (off by default), models show how people with similar computers rate them, and how everyone does, with how many ratings and a label when there are few.
  • Help translating Yggdrasil: CONTRIBUTING.md explains how to fix or review a language by pull request, and a Translation issue form takes reports of wrong text and requests for new languages. scripts/i18n.py status shows what a language still lacks, and scripts/i18n.py glossary <language> prints the words a language uses for Yggdrasil's main terms, read from its catalog (i18n/glossary.json).
  • The context gauge says how much memory the conversation window reserves on this computer, such as "about 1 GB", and that a longer window remembers more and uses more memory. Near the limit it says the oldest messages will be sent as a summary and suggests a new chat.
  • docs/design.md describes how the UI is built: color tokens with their contrast in both themes, type, spacing, the components and React helpers to reuse, the order of loading, failed, empty, and loaded states, keyboard patterns for dialogs, menus, and tabs, phone-width layout, accessibility rules and the CI check, writing and brand rules, and a checklist for new UI. The pull request template links it.
  • Email and calendar with your own accounts and an app password (no Google or Microsoft sign-in app). Email (IMAP and SMTP: Fastmail, iCloud, Proton Mail Bridge, Nextcloud, your own server, and Gmail or Outlook where app passwords are allowed) adds email.search, email.read (without marking messages read), and, asking first, email.draft, email.send (threaded replies, a copy in Sent), and email.archive. Nothing is deleted. Calendar (CalDAV) adds calendar.search and calendar.availability, and, asking first, calendar.create, calendar.update, and calendar.cancel, which keep attendees and alarms and never overwrite an event changed elsewhere. Connect them in Settings → Connected services.
  • Connect an OpenAI-compatible server, such as OpenAI, or vLLM or Ollama on another machine, in Settings → External server (advanced mode). Its models appear for a chat to choose, marked external, and chats with them are recorded in What left this computer. Auto never picks them, and offline profiles and chats using memories or knowledge marked This computer only refuse them.
  • Guided installation: asking for something Yggdrasil can't do yet but can install, such as "Make me an image of a Viking tree" before image generation is set up, is answered with what would be installed, its download size, and the computer it would run on. A card in the chat sets it up, shows the download, and then finishes the request. "Can you generate images?" offers the same card. The capability inventory lists these setups (setups).
  • Assistant language in Settings → Language: answers come in the language you write in (the default), in the App language, or in a language you choose, apart from the App language. Asking in a message, such as "answer in English" or "antworte auf Deutsch", always wins. The language is told apart on this computer; nothing is sent anywhere to find it out.
  • Automations choose the language their results are written in: the assistant language (the default), the App language, the language of the request, or a language you pick, under Advanced in the automation form. A language the request asks for, such as "answer in English", always wins.
  • pnpm lint in web/ fails on hard-coded text in the UI, so new text goes through the translation catalog.
  • BGE-M3 in the model catalog: a multilingual embedding model that lets knowledge and memory search find passages by meaning in more than 100 languages, such as a Spanish question finding an English document. When it is installed next to an English embedding model, it is the one used.
  • Knowledge sources record the languages they are written in, and the Knowledge page shows them. A source in another language than the App language says when a multilingual embedding model would help questions find it.
  • Memories are found across languages: with an embedding model installed, a memory saved in Spanish ("Mi proyecto usa Go") comes with a question in English ("What language does my project use?"), and the other way round. Each memory records the language it is written in, and the Memory page names it when it differs from the App language.
  • Models say how well they write each language: on the Models page, a card shows its level in your App language (Limited, Fair, Good, or Excellent) and how many other languages it has one for, and its details list every language with the confidence and where the level comes from. The API's catalog models carry languages.
  • Notifications in the API carry message, their title and body as translation keys with the values they need, and webhooks also get language. The client contract is 1.4.
  • Ratings by language: rating a model asks how well it worked for you in a language, your App language unless you choose another or not to say. A model card shows its community ratings by language, your App language first, and with community ratings on, Auto counts them when choosing a model for an answer in that language. The API's PUT /models/{id}/rating takes language, and /ratings/community returns each model's languages.
  • The web UI lays out right to left for right-to-left languages such as Arabic and Hebrew: the sidebar, drawers, menus, and arrows mirror, while code, commands, and logs stay left to right. A right-to-left pseudo-locale (ar-XB), chosen in Settings in advanced mode, shows layout that does not mirror.
  • scripts/i18n.py status counts only text that is probably untranslated: text meant to read the same as English, such as PDF or German Name, is listed in i18n/same-as-english.json, and CI says when a listed key is translated or gone.
  • Yggdrasil in German, Spanish, French, Italian, Brazilian Portuguese, Japanese, Korean, and Simplified and Traditional Chinese: the web UI, the desktop app's menus, and the iPhone app. The translations are machine made, and Settings says so under the language picker until each is reviewed. A system set to Chinese (Taiwan or Hong Kong) or Portuguese (Portugal) gets the closest one.
  • Pasted training examples can use the question and answer labels of these languages, such as P:/R: in Spanish or 问:/答: in Chinese.
  • Speech tools say which languages they work in (languages, and auto_detect for Whisper) in GET /tools/providers and in Diagnostics, and a speech call goes to a paired computer whose provider works in its language.
  • Images on this computer: image.generate makes an image from a description, and image.edit changes an image in the chat from an instruction, with stable-diffusion.cpp and FLUX.2 [klein] 4B. Image generation is set up once from the Tools page, which recommends a model for the computer's memory (5.2 GB, or 8.8 GB for more detail), downloads it with progress, and can stop, resume, and remove it (/api/v1/images/setup). Downloads are pinned and checked; prompts and images are not sent anywhere. It runs on macOS with Apple silicon, and on Linux and Windows on the CPU.
  • Chats accept PNG and JPEG images, which are shown in the chat and can be edited. Images the assistant makes are shown too.
  • An Images capability in profiles turns image generation on or off. Offline profiles keep it.
  • One command installs Yggdrasil and joins a computer to your network. yggctl join-token create now also prints an install-and-join command: install.sh on Linux and macOS, and install.ps1 on Windows, both attached to each release. They install the release for that computer, check it against the release's checksums, start Yggdrasil as a service (systemd, launchd, or a scheduled task), and join. A Yggdrasil that is already installed and running is left as it is.
  • yggctl join --name renames the computer as it joins, and --wait 60s waits for Yggdrasil to start first, for provisioning scripts. The Clustering guide has cloud-init and Ansible examples.
  • The Computers page can add a computer by command: Add by command makes a one-time join command to run on the new computer, for one with Yggdrasil, one to install it on, or Windows, with a copy button, a countdown, revoke, and the recent commands. It says when the computer has joined.
  • Diagnostics shows how much memory Yggdrasil itself is using and its background tasks, with the last day as a small chart. If memory grows steadily for hours, it says so and suggests exporting diagnostics to report it.
  • CI checks the web app for memory leaks by moving between every page many times, and a soak test (YGGDRASIL_SOAK=1 go test ./tests/soak) runs a throwaway daemon through hundreds of chats, cancellations, and event connections and fails if memory, goroutines, or open files keep growing.
  • Email, push, and webhook destinations can send a daily digest instead of each notification: choose Daily digest and a time. Errors still go out right away, and a digest due during quiet hours waits for them to end.
  • Join a computer to your Yggdrasil network with one command. On a computer in the network, yggctl join-token create prints a command; run it on the new computer, for example over SSH, and the two trust each other with no pairing prompt, mDNS, or configuration. The token lasts 15 minutes, works once, can be revoked, and never crosses the network: the command carries the first computer's fingerprint, so a different machine at that address is refused. yggctl network shows the computers this one trusts, and yggctl leave leaves the network.
  • Places and directions with OpenStreetMap: places.search finds a kind of place near somewhere (cafes, pharmacies, fuel, parks, and more) or a place by name, with addresses, opening hours, and distances in km and miles; places.details looks one up; maps.route gives driving, walking, or cycling directions; and maps.distance says how far and how long. They are part of the Internet capability. Each request is recorded in What left this computer as Maps and places, a repeat within the hour is answered from memory, and the services can be your own (places_geocoder_url, places_overpass_url, places_router_url).
  • Diagnostic bundles include a runtime summary and goroutine and heap profiles, so a report of growing memory can be diagnosed. They hold function names, counts, and sizes, no prompts or files.
  • YGGDRASIL_PPROF=127.0.0.1:6060 serves Go's live profiles on this computer for developers; only loopback addresses are accepted. See Troubleshooting: "Yggdrasil uses more and more memory".
  • Tools on other computers: image generation and speech run on whichever paired computer can run them, preferring one whose GPU does the work, so a laptop without image generation can ask for an image and a paired workstation makes it. The asking computer keeps the approval, the audit, and the files; the other computer runs the work and keeps nothing. The chat profile's computer policy applies, an unreachable computer is skipped for the next, and Stop stops the work there. Each job sent is recorded in What left this computer.
  • Diagnostics → Tools on each computer shows each computer's image and speech providers: ready, installing, failed, or not set up, and whether a GPU does the work (GET /api/v1/tools/providers).
  • Speech on this computer: speech.transcribe writes down what an attached audio file says (Whisper, fast or accurate), and speech.synthesize reads text aloud as an audio file (Piper). Answers have a Read aloud button (POST /api/v1/speech). Chats accept audio files (.wav, .mp3, .m4a, .aac, .ogg, .flac, .webm), which play in the chat. The first use installs speech from PyPI and downloads the model or voice from Hugging Face; audio and text are not sent anywhere.
  • A Speech capability in profiles turns transcription and reading aloud on or off. Offline profiles keep it.
  • Video on this computer: video.generate makes a clip of a few seconds from a description, or brings an image in the chat to life, with stable-diffusion.cpp and Wan 2.2 TI2V 5B. Video generation has its own setup on the Tools page and offers in chat (an 8.5 GB download; /api/v1/video/setup), runs on a paired computer that has it, and clips play in the chat. A Video capability turns it on or off in a profile.

Changed

  • CI checks every page for accessibility problems. It runs axe-core in a real browser over each page with demo data, in light and dark themes at desktop and phone widths, and fails on contrast, labeling, or structure problems (WCAG 2.2 AA). Run it locally with node scripts/screenshots/a11y.mjs; see docs/development.md.
  • Auto weighs community ratings and load. With community ratings on, a model people with computers like yours rate clearly better is chosen over a larger one. A model busy answering something else gives way to a similar idle one that's already loaded, and work goes to an idle computer before a busy one. The answer's steps say when ratings or load decided.
  • The separate Team orchestrator is gone. Profiles saved with it, and API requests that name it, move to the Team strategy with their roles and models; coordinator becomes planner. A profile's role models are now used in chat: before, the model chosen in the chat replaced them all, except in Team profiles.
  • Pull requests describe their changes in changes/unreleased/ instead of editing CHANGELOG.md, so they no longer conflict over the changelog. scripts/changelog.py checks the fragments in CI, previews the next release, and writes them into CHANGELOG.md when a release is cut.
  • Every Go package's tests fail if they leave goroutines running (internal/leakcheck), so leaks like these are caught before they ship.
  • The client contract is 1.2: answers can carry setup, an offer to install what the request needed.
  • The notes shown with an answer are in the App language: figures Yggdrasil could not confirm, an answer that claims a change no tool made, a stopped answer, a smaller model answering after another failed, and a small model answering from your files or knowledge. They were in English in every language.
  • The labels on an answer's sources ("Attached file", "Made in this chat", "Memory") are in the App language, including for apps that use the OpenAI-compatible API.
  • The steps listed with an answer are in the App language: what Yggdrasil looked up, read, ran, and checked, why Auto chose the model, and which model or computer answered after another failed. They were in English in every language.
  • Automations, notifications, Tools, connected services, and API Access are in the shared translation catalog (automations.json, notifications.json, tools.json, services.json, and apiAccess.json under i18n/locales/<language>/), so they follow the App language: schedules, notices, the notification bell and desktop notices, email/push/webhook delivery, tool sources, image generation setup, and API keys. Automations still read requests written in English, such as "every morning at 8:00 AM".
  • Chat's text is in the shared translation catalog (i18n/locales/<language>/chat.json), so it follows the App language: the composer, history, progress messages, tool permission prompts, errors, answer details, run details, and file chips. Plural forms, such as "Used 2 tools", come from each language's plural rules.
  • Errors from Yggdrasil show in the App language. The service sends each error with a stable code, such as MODEL_NOT_INSTALLED or MEMORY_LOOKS_SECRET, and the values its message needs, and the web UI shows the text for the code, in all ten languages; the English text stays under Details for Diagnostics. Chat recognizes errors such as a model running out of memory or a computer going offline by their code instead of by their English wording. Setting an invalid value in Settings now says which setting and value, instead of "Could not update settings."
  • The client contract is 1.3: a chat that fails while streaming sends event: error_code with the code, before event: error with the text, which older apps still read.
  • Numbers, dates, times, sizes, speeds, prices, and percentages follow the App language's region everywhere in the web UI: German shows 1.234 passages, 1,5 GB, 30.09.2026, and 45 %. "5 minutes ago" comes from the browser's own relative-time rules, so it reads naturally in every language.
  • The startup screen, the API key prompt, and the stories behind the Norse names, the mascot, and the logo show their text in the App language.
  • Knowledge, Train, Profiles, Orchestrators, and Diagnostics show their text in the App language. Counts on these pages use the language's plural forms, so the health page says "1 model installed" rather than "1 models installed".
  • Auto picks a model that writes the answer's language well: asked in Spanish, a model rated good in Spanish comes before a larger one rated fair. A model without language ratings is still used. When nothing installed writes the language well, the answer is still in it, with a note saying it may read less well and that better-rated models are on the Models page.
  • When a model stops responding or runs out of memory, chat says so in the App language.
  • Models, Computers, and Performance are in the shared translation catalog (models.json, computers.json, and performance.json under i18n/locales/<language>/), so they follow the App language, including the model notes Chat shows when a request needs a model that can use tools or see images. Counts such as "2 models running" and "3 computers" use each language's plural rules.
  • Notifications show in the App language: the bell writes each notice in the language you chose, and desktop notices, email, push, and webhooks are written in the App language when they are sent. Text a model or an automation wrote stays as it was written.
  • Chat messages, answers, and memories take the direction of their own text, so Arabic reads right to left in an English UI and English left to right in an Arabic one.
  • Run details (in advanced mode) show a failed run's error and the run's status in the App language. The English error text stays available on hover.
  • The client contract is 1.5: a run trace has error_code and error_details beside error, the same stable codes as other errors, so clients can show the error in the App language.
  • Run details (in advanced mode) show the strategy and effort in the App language: how the request was worked through, a specialized AI or automation that answered, and which model or computer answered after another failed.
  • Settings, onboarding, and Memory are in the shared translation catalog (settings.json, onboarding.json, and memory.json under i18n/locales/<language>/), so they follow the App language. Counts such as "18 cores" and "7 days" use each language's plural rules.
  • Automation schedules show times in the App language's clock, such as 18:30 in German.
  • Read aloud speaks the text's language: a German answer is read by a German voice, not the English one. There are voices for 25 languages; text in a language without one, such as Japanese, says so instead of being read in the wrong voice.
  • When a model runs out of memory because training is using this computer, chat says so in the App language, with about how long training has left and what to do. The explanation now shows even when the model's error was plain text, instead of only under Details.
  • "Can you …?" answers no longer count a tool that is on but cannot run on this computer, and say what would make it work, such as setting up image generation.
  • While a computer trains an AI, chats and other work go to a paired computer that has the model, and the answer's steps say which one answered. The Computers page marks a computer that is training. With no other computer that has the model, the chat is still answered here, as before.
  • Settings says plainly that your data stays private, and what is encrypted. A new Your data stays private card at the top of Privacy explains that chats, memories, knowledge, and files stay on this computer with no account, cloud, or telemetry. It also says Yggdrasil doesn't encrypt its own files, so turn on FileVault, BitLocker, or LUKS to keep them encrypted. Online services use HTTPS, but traffic between your own computers isn't encrypted yet. The user guide and docs/privacy.md answer "Is my data encrypted?" and "Are my chats private?", and so does the assistant when you ask it.
  • When a model can't answer, Yggdrasil tries another installed quantization of the same model before switching to a different model: a smaller one when it ran out of memory. The answer's steps say which version answered.
  • Shared community ratings go to https://ratings.toskar.ai by default, and the daily ratings summary comes from there. The ratings_url setting still overrides it.
  • The chat box is simpler. Model stays next to the message, and Profile, Run on, Effort, and Memory move into one Options button that shows what differs from the defaults, such as "General · Thorough · Memory off". On a phone the box is two rows instead of three rows of unlabeled menus. When Settings asks where chats run, Run on stays in view until you choose.
  • The sidebar's status footer uses plain words (Computers, Model placement, Knowledge) and is headed Status, so it no longer shares the System heading with the menu above it. The realm behind each line is in its tooltip.
  • Plainer wording where everyday users read it. Onboarding describes your hardware as "what this computer has for running AI models", not "hardware detected for local inference"; the Run on tooltip and the Knowledge page say what Yggdrasil does rather than naming Norn or Mimir in the sentence (the realm names still head each page); the sign-in screen and Diagnostics say "the Yggdrasil service" instead of "the daemon"; and a model's quantization is labeled "Compression (quantization)". In all 10 languages.
  • While a page loads, it shows placeholders in the shape of what's coming (model cards, list rows, chat bubbles) instead of a "Loading…" line, so the page doesn't jump when the content arrives. Searches and other work in progress keep their spinner. Models says "Checking hardware…" while it reads this computer's hardware, instead of "Hardware details unavailable".
  • .webm files are videos: they play as video in chat, and their sound can still be transcribed.

Removed

  • The Orchestrators placeholder page, which only said "coming soon" and wasn't in the sidebar. A link or bookmark to it opens Profiles & Orchestration, where each profile's way of working (its orchestration strategy) is chosen.

Fixed

  • Links in answers open. A model writing a page's address from memory often gets it wrong, so answers now link only to pages their sources gave: a guessed link goes back to the model to be replaced, and any left are named in a note under the answer, since they may not work. A site's home page is still fine. This applies wherever you chat, in the desktop app, the browser, and on a phone.
  • The assistant knows today's date and the time. Questions like "what's on this Friday?" or "how long until 5 pm?" used to depend on the model's guess; now every answer gets the date and time in your time zone, which the app takes from your browser or phone. Automations use the time zone their schedule runs in. Nothing extra leaves your computer.
  • The assistant knows Yggdrasil's own features and documentation. Ask how to do something in Yggdrasil, such as scheduling an automation, connecting a computer, or adding knowledge, and it answers from the user guide that ships with the app, naming the screens and buttons to use. Before, a model guessed. Questions about anything else are unchanged.
  • The web UI in Brazilian Portuguese and Simplified and Traditional Chinese no longer falls back to English.
  • An automation's amount accepts a decimal comma, such as 19,99.
  • Yggdrasil Desktop: answer sources and other links that leave the app now open in the default browser. Signing in to an MCP tool source now works too: it opens in the browser and returns to the daemon's address, where before it was sent back to the app's own wails:// address, which no browser can reach. Saving a chat file, an exported GGUF model, or a training sample now asks where to save and writes the file; a large model streams straight to disk. The desktop app's web view can't open windows or download files, so these go through the desktop shell (Yggdrasil Desktop 1.4 or later). In a browser, nothing changes.
  • Yggdrasil Desktop on Windows: live progress, tool approval prompts, notifications, and chat replies as they are written now appear. Wails hands the app's web view a proxied response only once it ends, and the daemon's event stream never ends, so none of it arrived; replies showed all at once. The desktop shell now reads the event stream and relays each event to the page, and replies take their text from that relay (Yggdrasil Desktop with the event relay). The desktop app uses the relay on every platform; in a browser, nothing changes.
  • The context gauge counts everything in the system prompt, including your personalization, memories, and guide excerpts, which it left out when estimating.
  • Chats with an external server show the real prompt size: Yggdrasil asks the server for its token counts at the end of each reply.
  • When no tokenizer is available, the estimate is closer for Chinese, Japanese, and Korean (about a token per character, not a quarter) and for code and JSON.
  • The chat's context gauge uses the window the model is actually running with, read from llama-server, instead of a guess from the catalog that could differ from it. Switching back to a chat shows its last reading instead of an empty gauge.
  • Connected services scrub only secret values, such as a token, from results. Every stored value of six or more characters was scrubbed, so a result mentioning the service's own address, such as a Home Assistant URL, read "[redacted]".
  • The external-openai runtime could not be configured, so it always reported "not configured" though the documentation described it as working.
  • Looking for nearby computers no longer risks returning a half-written list. The search now waits until every answer it heard has been read before handing back the results, instead of guessing with a fixed pause.
  • The daemon starts again on builds where two database migrations shared number 031: knowledge source languages are now migration 032.
  • Yggdrasil starts again on a build with both knowledge languages and shared model runtime observations. The two shipped database migrations with the same number, so the database refused to open; knowledge languages now has its own number.
  • Cancelling training on a paired computer just as it starts now stops it there. Before, a cancel that arrived while the examples were still being sent could leave the paired computer training until its daemon restarted, holding its training slot. A cancel now waits for the paired computer to answer (up to 2 minutes), then stops the run there and removes its files.
  • The sidebar header no longer runs the status chip ("No model", "Ready") into the notification bell. The status sits under the Yggdrasil title, and a label too long for the space is shortened, with the full hint on hover, in every language and in right-to-left layouts.
  • Web page reads (internet.open) no longer open addresses on your own computer or local network, such as Yggdrasil's own API, a router's admin page, or a cloud metadata service. A web page could otherwise lead the assistant there and read the result. Every address a site's name resolves to is checked, again on each connection, so a redirect or a changing name cannot get around it.
  • Checking email no longer leaves a background task behind for each IMAP connection. A daemon that checked mail on a schedule grew without limit.
  • The browser tool's idle-session cleanup stops as soon as the last browser closes, including when Yggdrasil quits, instead of up to a minute later.
  • A few words left in English are translated: the API key placeholder in Japanese, and the LoRA rank and epoch fields, the loopback address note, and a suggested prompt in Traditional Chinese.
  • The .deb and .rpm packages of a pre-release, such as 1.4.0-beta.1, have a valid version again; it had the build machine's home folder in it, so apt and dnf refused them.
  • Renaming a computer in Settings shows the new name to other computers right away, instead of after a restart.
  • A computer that joins a network under a name already taken there takes the name the network gave it, such as "worker-01-2", so both computers show the same name.
  • Renaming a computer while discovery settings change no longer leaves an old network announcement running.
  • Asking the assistant how to install or add MCP (or "MPC") now gets the actual steps: Tools → Add tools and its four ways in, or, for using Yggdrasil from Claude Desktop, Cursor, or VS Code, where to find what to paste. Before, a model guessed, often telling people to edit another app's config file.
  • Paired computers' fingerprints are stored in the same form join commands show; existing ones are updated when Yggdrasil starts.
  • Stopping an MCP tool source ends the helpers it started, such as the node process npx runs, even when the server quits on its own. They used to keep running each time a source stopped.
  • After a model ran out of memory, the fallback could pick a larger version of the same model; it no longer does.
  • Text is easier to read. Faint labels, status chips, and accent colors now meet the WCAG AA contrast standard (4.5:1) on every surface, in both light and dark themes. Light-theme chips such as Ready and Offline were the hardest to read before.
  • A page that runs into a problem now says so and offers Try again, and the rest of the app keeps working. Before, one page's error blanked the whole window.
  • Each page names itself in the window title, such as "Models · Yggdrasil", so browser tabs, history, and screen readers say where you are.
  • Screen readers get a clearer page: the sidebar is a header, the Main navigation, and a footer; empty states and computer cards use the right heading level; and the Default profile menu in Settings is labeled.
  • Yggdrasil works fully from the keyboard. Dialogs (delete a chat, approve a tool, rate a model, a tight fit, local network access) and the chat history and phone menu drawers keep focus inside while open, close with Escape, and return focus to what opened them. Dialogs start on the safe choice, so Enter alone never deletes a chat or approves a tool, and Escape on a tool request denies it.
  • Tabs (Models, Performance, Profiles, adding a tool source, sharing with apps, join commands), star ratings, and the menus on chats, models, and profiles follow the standard arrow-key patterns: one Tab stop per group, arrows to move, Home and End, and Escape to close a menu.
  • The chat box and the chat-list search show when they have focus. The setup choices in onboarding say which one is selected.
  • A page that can't load its data says so, with Try again and a link to Diagnostics, instead of looking empty. Before, a failed request showed "No computers", "No model installed", or "No tool sources yet", as if they were gone, or kept saying Loading. This covers Chat, Automations, Models, Train, Knowledge, Memory, Computers, Performance, Profiles, and Tools.
  • The sidebar's status no longer says "No model" or "No computers" before it knows; while those are loading or can't be read, it doesn't claim either. A failed request is retried once instead of three times, so its error shows in about a second.
  • Yggdrasil works in a phone's browser and in a narrow window. Below tablet width the sidebar used to stay full size and leave pages about 140 pixels, one word per line. Now it is a menu that slides in from a button at the top, and pages use the full width with a narrower margin.
  • On a phone, the chat box no longer covers the conversation. In a chat it starts at one line and grows as you type, up to 30% of the screen; the empty start screen keeps its large box. The History and New chat buttons stay on one line instead of wrapping.
  • The App Store and Google Play phone screenshots show the app as it is on a phone, with the top bar and menu, instead of a layout made only for the screenshots.

Security

  • Pairing two computers is signed end to end. Each computer proves it holds its key, and only the computer that was asked can finish a pairing, with the code that was shown. Update Yggdrasil on both computers before pairing them or using them together.
  • Pairing codes are limited: repeated wrong codes or failed answers end the pairing, and you start again with a new code.
  • Pending pairing requests are no longer listed to other computers on the network.
  • Requests between paired computers name the computer they are for and are good only once.
  • Pairing never replaces the key of a computer that is already paired. To pair a computer again with a new key, such as after reinstalling it, remove it on the Computers page first.

1.4.0

October 2, 2026

Downloads and release notes ↗Docs for 1.4.0 →

Stable release of 1.4.0. It contains everything in 1.4.0-beta.1, the AI experience platform and the remaining Train Your Own AI items, plus the changes below. Existing API routes, configuration, and data are compatible with 1.3: the changes add routes, optional fields, and database tables, and migrations run automatically. NVIDIA (CUDA) training, PostgreSQL and MySQL knowledge sources, and the Mac App Store sandbox have not been tested on that hardware or in that build. Binaries and the apt repository are not signed.

Added

  • Documentation for 1.4.0. New Configuration (data directory layout, every config.json key, environment variable, and setting) and Command line (yggdrasil-daemon flags, the Linux systemd service, and every yggctl command) pages. The API reference now lists every route, grouped by area, and every event type. The user guide adds sections on memory, knowledge, tools, notifications, profiles, privacy, and diagnostics.

Changed

  • api/openapi.yaml now describes every route: the control plane, the OpenAI-compatible API, and the MCP server, with request bodies, responses, the error shape, and bearer authentication. It previously covered 51 of them. A test fails when the daemon serves a route the spec does not describe, or the spec describes one the daemon does not serve.
  • Rewrote the architecture, privacy, and troubleshooting pages for 1.4.0, and updated tools, capabilities, compatibility, and the README. Removed statements that no longer matched the code, such as where retrieved knowledge goes in a prompt and which /chat fields exist.
  • Each release carries its own screenshots. After a release is published, the Screenshots workflow captures the README stills and the walkthrough from that release's interface and attaches them as screenshot-<name> files, which yggdrasil.yeix.io shows. Started by hand with a tag, it attaches them to an existing release.

Fixed

  • Chat answers keep their indentation. Nested lists stay nested and code blocks keep their indentation; only extra spaces in the middle of a line are collapsed.
  • The Screenshots workflow passes again. Screenshot validation accepts the lowercase opaque value that ImageMagick 6 on Ubuntu prints, a screen that does not become ready is loaded once more before the run fails, and a failure now prints the page's errors and keeps what the page showed.

1.4.0-beta.1

October 1, 2026Pre-release

Downloads and release notes ↗Docs for 1.4.0-beta.1 →

Beta pre-release of 1.4.0. It adds the AI experience platform and the remaining Train Your Own AI items. Existing API routes, configuration, and data are compatible: the changes add routes, optional fields, and database tables, and migrations run automatically. NVIDIA (CUDA) training, PostgreSQL and MySQL knowledge sources, and the Mac App Store sandbox have not been tested on that hardware or in that build. Binaries and the apt repository are not signed.

Added

  • Sandboxed Mac App Store builds can train on a paired computer. A sandboxed copy of Yggdrasil no longer tries to download Python; it says training can't run on this computer and chooses a paired computer running Yggdrasil Core. A store build can also ship the training and text-recognition environments beside the daemon, and yggdrasil-daemon -python-envs lists what to bundle.
  • Versioned client contract. Events, run traces, and answer metadata (citations, steps, files) carry a contract version (1.0), every API response has a Yggdrasil-Contract header, and /api/v1/version describes the contract. Fields are only ever added within a major version; a test fails if one is removed or renamed. An app built for another major version gets a clear 426 that says which side to update.
  • Caching with declared policies. Every cache says what it keeps, how long, what clears it, where it applies, and how private it is; credentials are never cached. Repeat web searches and page reads within minutes are answered from memory, so nothing leaves the computer again. The capability inventory is cached and refreshed when models, computers, or tools change, and Hugging Face searches use the same cache. Diagnostics lists the caches in advanced mode, run details show cache hits, and deleting run records clears personal caches. Routes are under /api/v1/caches.
  • Train on NVIDIA GPUs. A computer with an NVIDIA GPU can now train specialized AIs with PyTorch, using LoRA, or QLoRA when the GPU's memory is tight. Training fit uses the GPU's own memory, and a Mac can send training to a paired NVIDIA PC. The first run installs PyTorch with the CUDA libraries for the computer's driver (about 4 GB).
  • Knowledge from databases and web APIs. Connect a SELECT query on a SQLite file, PostgreSQL, or MySQL, or a URL that returns JSON, CSV, or text, and each row or item becomes a passage. Yggdrasil only reads: queries run read-only. Data older than the chosen interval (5 minutes to a day) is fetched again when a question uses it, and if a fetch fails the last data keeps answering. Passwords and tokens are stored apart from the database and never shown again.
  • Export a specialized AI as one GGUF file. The Deploy step merges the trained revision into its base model, so LM Studio, Ollama, llama.cpp, and other GGUF tools can run it without Yggdrasil. The file is a little larger than the base model, and the AI's instructions are shown to copy as the system prompt. A notification says when a large export is ready.
  • Quality test set. Ten representative requests, each with the behavior it must have, run against the stub model on every change and against real models with make quality-real or a weekly self-hosted workflow. The behavior checked includes a simple question staying direct, a price question using knowledge, a risky command asking first, a long conversation remembering an early fact, a current question being looked up, and a request with parts being planned. Running it against Llama 3.2 1B led to three fixes. Plain questions no longer offer tools, which the small model misused. Short capability questions are answered from the inventory. An answer that claims a change no tool made is now called out.
  • Scanned PDFs in Knowledge. Pages without a text layer are read with text recognition, so scanned manuals, warranties, and price sheets become searchable and are cited by page. The first scanned PDF installs the recognizer (about 110 MB) on this computer; nothing is sent elsewhere. A scanned PDF attached to a chat explains how to connect it on the Knowledge page instead.
  • Capability inventory. Yggdrasil keeps track of which models, computers, tools (built in, connected, and MCP), connected services, providers, and files exist right now, and what they let it do. Ask "Can you generate an image?" or "Which computer can run Qwen 2.5 14B?" and the answer comes from that inventory instead of a guess. Diagnostics lists every ability, with how it works or what would add it. Route: /api/v1/capabilities.
  • Structured results. /v1/chat/completions supports response_format (json_object and json_schema). Answers are checked against the schema, safely repaired, and asked for once more if needed; JSON that still does not fit gets a 422 that lists the problems. Tool arguments are checked and repaired before a tool runs. Price and significance automations read their result's JSON with the same repairs, so "$1,299" counts as a price, and ask the model once when it is missing.
  • Profiles & Orchestration. In advanced mode, a profile has an Orchestration section that sets its reasoning level and planning, as well as workers, parallelism, verification, tool calls, memory, context budget, fallback, and a time limit. Blank keeps each default.
  • Run details. Every chat, API request, and automation is traced. In advanced mode, each answer has "Run details", showing the strategy, effort, models with their computer, load time, time to first token, tokens per second, and cached tokens, as well as tools with timings, plan workers, verification passes, retries, context size, and latency. Routes are /api/v1/runs.
  • What left this computer. Settings lists every web search, page read, paired computer, external server, and connected service that a chat, automation, API request, or training run sent data to, with a 30-day summary. Memories and knowledge sources can be marked "This computer only"; chats that use them run here even when a paired computer would otherwise answer. Run records (stored prompts and tool results) are kept for 30 days by default, with a choice of 7 days to keeping them, and can be deleted at once. Routes are /api/v1/egress and /api/v1/privacy.
  • MCP, both ways. Tools → Add tools adds MCP servers as tool sources: pick one from the gallery (Folders, Browser, Notion, Linear, Jira & Confluence, Sentry, GitHub, Context7, DeepWiki, SQLite, PostgreSQL, Brave Search, and more) and answer a question or two; paste any app's settings, a web address, or a command line; or bring the servers already set up in Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Gemini CLI, or LM Studio. Services with a sign-in open one in a browser window (OAuth with PKCE and app registration). Their tools work in chat and automations like connected services: reading runs, changes ask first, results are untrusted data, and secrets stay in the secrets directory and are scrubbed from results. Servers on this computer start when a tool is needed and stop when idle, get only a safe environment, and say plainly what to install when Node.js or uv is missing. Each source has a log, per-tool Ask first and on/off, ready-made prompts that start a chat, resources, and an opt-in for servers to ask the AI for help. Other apps can use Yggdrasil too: /mcp offers ask_local_ai, list_local_models, and search_my_knowledge within an API key's permissions, yggctl mcp bridges apps that start a program, and API Access gives the settings to copy for each app. Routes are under /api/v1/mcp. See MCP.
  • The API gets the same assistant. /v1/chat/completions uses the whole conversation, including the system prompt, not only the last message. reasoning_effort sets the effort. An optional yggdrasil object opts into memory and connected knowledge, narrows tools, chooses effort and placement, and streams progress and tool activity. Answers carry their sources and steps. Each API key has permissions on the API Access page (memory, knowledge, tools, placement) that requests can narrow but never widen. API requests no longer use your memories unless they ask.
  • Personalization. In Settings, choose answer length, tone, format, and units, and add a note about yourself and how you like answers. It applies to every chat, automation, and API request. It is kept apart from permissions: a preference or memory that tries to grant one ("you can always push without asking") is refused, with a pointer to Tool permissions.
  • Knowledge search by meaning. Install an embedding model, such as Nomic Embed Text v1.5 (new in the catalog, 146 MB), and Mimir finds passages that answer a question even when they use different words: "warranty" finds your guarantee policy. Word matches still count, and passages both searches find rank first. Passages are embedded in the background and only again when their text changes; a chat, an automation, or training goes first. The Knowledge page says when a source is searchable by meaning. An installed reranker model reorders the best passages. Without an embedding model, search works as before.
  • Connected services: GitHub and Home Assistant. Connect them in Settings with a token. The AI can then search and read issues and pull requests, comment (after asking), check lights and sensors, and control devices (after asking). Tokens stay on this computer outside the database, are never shown again or given to the AI, and are scrubbed from anything a service returns. Settings explains the narrowest access to grant. Routes are under /api/v1/connectors.
  • Auto knows your specialized AIs. A question about what one was trained for (or one that names it) goes to that AI, and "What I did" says why. Questions that need the web, your files, or code still go to a general model, because specialized AIs answer without tools. An AI whose trained adapter is missing from this computer is skipped and explains itself.
  • Embedding, reranker, and classifier models are recognized as supporting models. They are labeled on the Models page, and they are left out of the chat and automation model menus. Auto and fallback never pick them to answer.
  • Sharing the computer. Chat comes first, then automations, then benchmarks, then training. An automation waits for your chat to finish instead of loading a model alongside it, a benchmark no longer unloads the model a chat is using, and training frees memory only once nothing else is running. Waiting work says what it is waiting for. A chat during training is still answered, with a note that training is using the computer and about how long is left. An automation that runs out of memory twice in a row is paused and tells you why, instead of failing every day.
  • Notifications. The bell next to the Yggdrasil name shows unread notices: finished and failed automations, tools an automation skipped, finished or failed model downloads, deployed AIs, and newly paired computers. Click one to go to it, mark all read, or dismiss. Notices are kept by the daemon, so ones that arrived while the window was closed are waiting. Desktop notices are still posted for automations, and each delivery is recorded. Routes are under /api/v1/notifications.
  • Automations use what chat uses: Auto can pick the model for each run, and memories and connected knowledge apply. An automation can notify only when it fails. Tools are approved when the automation is saved, including tools that change things, which the form lists apart. A run that reaches a tool you did not approve skips it, finishes, and tells you which tool to approve, instead of failing.
  • Redrawn Yggdrasil mark: vector source, interface colors, small-size version, theme-aware favicon. The sources are in docs/brand/logo/, and make icons renders the Linux icons from them.
  • Ratatoskr, the Yggdrasil mascot. He appears at the moments that matter: thinking while a reply is written, delivering work to a paired computer, celebrating a finished download or deploy, dropping his acorn on an error, and asleep when no model is loaded. He is a still frame when the system asks for reduced motion.
  • Only the tools a request needs. A weather question is offered web search, not the terminal or Git; a request about a file or a commit gets file or Git tools. A tool that was not offered is refused, so the model cannot reach beyond it. Tool calls have time limits and report why they failed. web.search, files.read, shell.run, and similar capability names reach the built-in tools, and a call a small model writes out as text is still run.
  • Effort in chat: Auto, Fast, Balanced, or Thorough. Fast answers in one go; Thorough reads more pages, uses the largest model that fits, and checks figures twice. Auto keeps quick questions fast and gives questions about your data, and requests with several parts, more care. The choice is remembered.
  • Stop means stop. Stop ends every model call, tool, plan step, approval, and paired computer working on the reply, from any window or through POST /api/v1/chat/stop, and keeps what was already written, marked as stopped.
  • Big requests are worked through in parts. "Compare Ollama, llama.cpp and MLX" looks each one up on the web side by side; "find three NAS drives, then compare price per TB, then make a spreadsheet" runs step by step, each step building on the last. A checklist shows the parts while they run, and one answer (or file) comes from all of them.
  • Answers are checked before you see them. Calculations are recomputed, and figures in answers that use your files, knowledge, or web results must appear in the lines about the same thing. A wrong figure is sent back to the model to fix once. Anything still unconfirmed is called out under the answer ("could not confirm 20 in the sources"). An answer that only describes tools, instead of answering, is asked for again without tools.
  • Small-model notes. On the Models page, models under 4B parameters say they can mix up facts and numbers from your files and knowledge, and suggest a larger model that fits the computer. In chat, an answer from a small model that used your files or knowledge carries the same note. Auto prefers a larger model for questions about your data.
  • Files in and out of chat. Attach documents, spreadsheets, PDFs, and code files with the paperclip, by dropping them on the message box, or by pasting. The answer cites the file, and later questions in the chat can still use it, including files the assistant made ("add a column to that spreadsheet"). Ask for a file ("make a spreadsheet of these prices") and the answer comes with a download. Spreadsheets are real .xlsx workbooks. Files stay on this computer, under artifacts/ in the data directory, and are deleted with their chat. Routes are under /api/v1/artifacts.
  • Each page shows its Norse name above the title (Ratatoskr for Chat, Mimir for Knowledge, Bifrost for Computers, and so on), and each tab in the sidebar has its Elder Futhark rune. Tab names are unchanged. Click a Norse name, the mascot, or the logo for a short lore entry: who it is in the myths and what it is in Yggdrasil.
  • Auto model. New chats use Auto, which picks an installed model for each message. Coding questions go to a coding model, questions about current information to a model that can use tools, and quick questions to a fast model that is already loaded when there is one. The answer's "What I did" says which model answered and why. auto is also a model in /v1/models.
  • Current questions are looked up first. When a question needs current information (weather, news, prices, scores, links) and web search is allowed, Yggdrasil searches the web and reads the best page before the model answers, so small models answer from the page instead of guessing or picking the wrong tool.
  • Quiet recovery. If a model fails before it answers, another installed model answers instead, and the answer says so, with a note when the model that answered is noticeably smaller. Auto skips a model that failed in the last 10 minutes.
  • Memory. Say "Remember that…" in any chat, and Yggdrasil keeps it across chats, restarts, and model changes; "Forget…" and "What do you remember?" work too. Answers that used a memory list it as a source. The Memory page lists memories by category, where you can add, edit, pause, or delete them, and turn memory off. Each chat has a Memory on/off switch. Passwords, keys, and card numbers are not saved. Routes are under /api/v1/memory.
  • Long conversations keep working. When a chat's history passes half of the model's window, older messages are summarized after the reply and the summary takes their place; the messages stay saved. The context meter shows how many messages the summary covers.
  • Chat answers show their sources (web pages, knowledge passages, and files) and a "What I did" summary of searches, pages read, and knowledge used. Errors are explained in plain language with a next step and a retry; technical detail is under Details. Knowledge lookups show progress while the answer is prepared.
  • Mimir connected knowledge. Connect a file, a folder, pasted content, an Excel workbook (each sheet is a table), or a PDF with a text layer (each page is cited); chat adds the passages that match each question. File and folder sources reindex when the files change. A profile lists knowledge_sources. Routes are under /api/v1/knowledge.
  • Train your own AI. The Train page builds a specialized AI from a base model, examples, instructions, and connected knowledge. Yggdrasil recommends Training, Knowledge, or Both for each piece of material, flags weak examples, estimates training fit separately from inference fit, trains a LoRA or QLoRA adapter with MLX on Apple Silicon, and compares base and specialized answers before deployment. A deployed AI is the model sai:<name> in chat and /v1/chat/completions. The first training run installs a private Python environment under runtimes/python and downloads the base model's training weights from Hugging Face. Norn can train on a paired computer that has more memory, and the Review step lets you pick the computer; the adapter returns to the computer that owns the AI. The Train page and the Profiles editor attach existing knowledge sources, files, and folders. "Try an example" sets up a sample tire shop assistant with notes on each step, and the Material step shows sample files in each format.
  • yggctl completion <bash|zsh|fish> prints a completion script for yggctl commands, automations subcommands, and their flags. Homebrew, the deb and rpm packages, and the macOS and Linux archives install or include the scripts.

Changed

  • Context budgets count tokens with the running model's own tokenizer. Which earlier messages fit, when a long conversation is summarized, and the context gauge's sections use llama-server's count instead of guessing four characters per token, a guess that is often well off for code and for languages other than English. Counts are cached for an hour and cleared with run records. When the model is not running on this computer, the estimate is used and the gauge still shows "~".
  • Connected knowledge and retrieved content reach the model as labelled data in the user turn, not in the system prompt, and the model is told not to follow instructions inside it. After a turn reads untrusted content, a tool that changes something (write, terminal, Git commit or push) asks first even when the profile allows it. Knowledge search drops passages that score far below the best match, and table columns such as in_stock read as "in stock".

Fixed

  • The context gauge undercounted every turn after the first: it counted only the prompt tokens llama-server processed, not those it reused from its cache. It now shows the whole prompt.
  • A daemon started with --data-dir stays in that directory when its config.json does not list data_dir. Before, it used the default data directory, so a second or test daemon could open your real database.
  • The database refuses to start with two migrations of the same number, instead of silently skipping one.
  • The chat page no longer reopens its event stream on almost every render, which dropped events such as a plan's checklist. The first message of a new chat no longer disappears while the reply is being written.
  • A model whose llama-server exits while loading, for example a damaged file, now fails at once instead of after a two-minute wait.
  • More current-information questions are recognized (news, scores, prices, exchange rates, "near me"), and cues match whole words only.
  • Chat starts faster when a paired computer is offline. Peer health is checked in the background every 10 seconds and reused for 20; a check that has to run during a turn waits at most 1.5 seconds. Before, each turn waited the full timeout for an offline peer.
  • A fast reply, such as a memory confirmation, no longer shows twice.

1.3.1

September 29, 2026

Downloads and release notes ↗Docs for 1.3.1 →

Patch release. The API is unchanged. Binaries and the apt repository are not signed.

Fixed

  • A saved schedule turns on “Keep running in background,” so closing the window does not stop the daemon that runs it.

1.3.0

September 29, 2026

Downloads and release notes ↗Docs for 1.3.0 →

Minor release. The API, CLI, and database migration are backward-compatible. Binaries and the apt repository are not signed.

Added

  • Scheduled automations. The daemon runs a saved prompt on a one-time, daily, weekly, or interval schedule, keeps a history of each occurrence, retries timeouts and connection failures, and can post an operating-system notice when the result matches the notification rule. Read-only tools can run while the window is closed. Automations are available in the desktop UI and through yggctl automations. GET /api/v1/automations is new.

1.2.1

September 28, 2026

Downloads and release notes ↗Docs for 1.2.1 →

Patch release. The API is unchanged. Binaries and the apt repository are not signed.

Added

  • Branching and release strategy guide in docs/development/branching-and-release-strategy.md.

Changed

  • Flat logo in the web UI, favicons, home-screen icons, and Linux package icons.
  • make start builds the UI and daemon and runs them. make help lists targets.

1.2.0

September 28, 2026

Downloads and release notes ↗Docs for 1.2.0 →

First stable release. It follows v1.2.0-beta.3. Binaries and the apt repository are not signed.

Added

  • Unsigned Windows amd64 headless archive on the GitHub Release.
  • Homebrew formula written and merged from the release workflow.
  • Coverage badge, golangci-lint, and ESLint in CI.
  • Repository guides for contributors, security reports, privacy, architecture, compatibility, and issue forms.
  • Feature specifications and the distributed-inference research brief. Those documents are plans, not shipped behavior.

Changed

  • Web UI uses Tailwind 4.
  • The README leads with the local-AI goal, the demo, and packaged install paths.

Fixed

  • llama.cpp health errors include the install location when the runtime binary is missing.

Earlier releases are in the full changelog and the GitHub releases.