Skip to content

Documentation / 1.14.1

Tools, permissions, and connected services

Choose what Toskar may do besides answering, such as searching the web, reading and writing files, or running code, and when it asks you first.

Profiles control which tools a model may use. Turning on a capability allows its tools. In advanced mode, each tool has a permission: Always allow runs it, Ask each time and Ask once per session ask in chat first, and Deny never runs it. Settings sets the defaults for the terminal, file writes, and Git. A tool turned off on the Tools screen cannot run anywhere.

The Tools page has three tabs. Your tools lists every tool, built in or added, with search and filters, and turning one off keeps it out of every chat. Add more connects services and adds tool sources, and Images and video sets up making images and clips. How it works explains that tools are built in or added, and that each profile chooses which tools its chats may use.

Internet searches DuckDuckGo and reads public pages. Files finds, reads, and writes workspace files. Terminal runs commands. Git inspects a repository, commits, and pushes. Files you can download are made with files.create, which writes only to Toskar's own file store. After an answer reads a web page or a file, a tool that changes something asks first even when the profile allows it, because retrieved content is data, not instructions. Text on a page or in a file that gives orders to an AI, such as “ignore your instructions”, is left out before the model reads it.

Each message is offered only the tools it needs: web search for current questions, files for a file or folder, the terminal for “run” or “install”, Git for “commit”. A plain question gets no tools. The model cannot call a tool it was not offered. How many tool calls a message may make depends on effort: 3 on Fast, 10 on Balanced, and 16 on Thorough.

Connected services add tools for GitHub, Home Assistant, email, and your calendar. Connect them on the Tools page, on the Add more tab under Connected services, with a token. For email and calendar, give your mail server, CalDAV address, and an app password from your provider (Fastmail, iCloud, Proton Mail Bridge, Nextcloud, or your own server; Gmail and Outlook if they allow app passwords). Then ask "Anything from Sam this week?" or "Am I free Thursday afternoon?"; Toskar asks before sending, archiving, or changing an event; Toskar checks it, stores it apart from the database, and adds it only when a tool runs, so it is never shown to the model. Reading is allowed and changes ask first unless a profile says otherwise.

Flight status answers "What gate is AA3595 leaving from?" or "Is UA 912 on time?" with the gate, terminal, scheduled and expected times, how late it is, and the baggage belt, from AeroDataBox. It needs your own AeroDataBox key: sign up on API.market or RapidAPI (the free tier covers a few look-ups a day), then connect Flight status under Connected services and paste the key. A RapidAPI key also needs the API address https://aerodatabox.p.rapidapi.com. Each look-up sends the flight's number and date to AeroDataBox, and What left this computer lists it. Without a key, flight questions use web search, which rarely knows the gate.

Reolink camera connects to a Reolink camera or doorbell on your network, straight through the camera, with no hub software or cloud. Connect it under Connected services with the camera's address, a user, and its password. A doorbell paired to a Home Hub uses the hub's address and its channel there, starting from 0. An automation can then look at the camera as it runs, or start when the camera sees a person or someone rings the doorbell. It can also speak through the doorbell: install Reolink's reolink-cli on this computer, and Toskar uses it with the connection's own address, channel and login, with nothing more to set up. Then an Admin's automation can choose Say the answer at the door, the voice it speaks in, and quiet hours when it only notifies. Toskar says nothing when the picture shows nobody to talk to, unless someone rang the doorbell. You're told who came and what was said at the door. Choose Talk with the visitor and Toskar greets them the moment it sees them, without looking at the picture first, so someone passing by hears the greeting too. Then it listens through the doorbell's microphone and takes turns: it finds out who came and why, up to six answers, and never says whether anyone is home or offers to open the door. It doesn't listen while it speaks. After saying goodbye it still listens for a minute, in case the visitor has more to say. When the visit ends, you're notified with a short summary and what was said. Hearing the doorbell needs ffmpeg on this computer as well; on a Mac, brew install ffmpeg. Toskar starts reolink-cli's gateway, at 127.0.0.1:9000, if it isn't running, and restarts it when it stops reaching the doorbell.

MCP servers add more tools. On the Tools page, the Add more tab's Add tools offers a gallery of servers, servers already set up in your other apps, pasted configuration, or a custom server. Tool sources lists each server, its tools, its log, and a Sign in action for servers that sign in with a browser. Other AI apps can use Toskar too: API Access → Use Toskar in other AI apps shows what to paste into Claude Desktop, Claude Code, Cursor, VS Code, and other apps.

Ask for a Word document, a PDF, or a spreadsheet and Toskar makes it, ready to download: documents and PDFs from formatted text with headings, lists, and tables, and spreadsheets with several sheets and formulas. Attach a spreadsheet and ask about it, such as "What is the total by region?", and Toskar reads the whole file, not just the part that fits in the prompt.

In advanced mode, the Tools page shows each tool's level (1 is low risk on this computer, 4 runs commands or code), what it returns, its time limit, and whether it is ready. It also lists recent calls: what each was about, how it was allowed, and how it ended.

Ask Toskar to calculate, analyze, or chart something, such as "Plot monthly sales from sales.xlsx", and it can write and run Python in a sandbox: it has no internet, sees only the files from the chat it is given, and stops after 90 seconds. Charts and files it makes are attached to the answer. It asks before running unless the profile allows it, and the first run installs Python with numpy, pandas, and matplotlib. On macOS it uses the built-in sandbox; on Linux it needs bubblewrap; on Windows it is not available yet.

For pages that need clicking or typing, Toskar can use a browser: "Go to the library's site and find this week's events". It runs the Chrome, Edge, Chromium, or Brave already on your computer in a fresh, separate profile, never your own sign-ins, and it can't reach anything on your own network. It asks before clicking, typing, or downloading, never types passwords or payment details, and can attach a screenshot so you see what it sees.

Ask about places, such as "Coffee near the Juneau library", "How far is the airport from downtown?", or "Walking directions from the ferry terminal to the capitol", and Toskar looks them up with OpenStreetMap: names, addresses, opening hours, distances in your units, and turn-by-turn directions. Each lookup appears in What left this computer. To ask "Coffee near me?" or "What's the weather?" without naming a place, type your town under Settings → Personalization → Your location. It's used only for questions about nearby places, weather, and directions, and the answer says when it used it.

Attach a voice memo or recording and ask what it says, and Toskar transcribes it on this computer. Answers have a Read aloud button, and you can ask for something to be read aloud as an audio file. Audio plays right in the chat. The first time, Toskar installs speech and downloads a voice or speech model; after that it works offline.

Toskar can make images on this computer, and change them: ask for "a watercolor of a lighthouse at dusk", or attach a photo and ask to "make it night". Set it up once, from the chat when you first ask for an image or from the Tools page's Images and video tab: it recommends a model for your computer (a 5.2 GB download for most) and shows the download as it goes. After that, images are made offline, and nothing is sent anywhere. It works on Macs with Apple silicon; on Linux and Windows it runs on the processor, which is much slower. With a paired computer that has it set up, such as a Mac Studio, a laptop without it can ask for images too: the work runs there, and Diagnostics shows which computer can run what.

Toskar can make short video clips too: ask for "a short video of waves rolling onto a beach", or attach a photo and ask to bring it to life. Set it up once, from the chat or the Tools page's Images and video tab (an 8.5 GB download, for computers with 16 GB of memory or more). A clip of a few seconds takes several minutes, and it plays right in the chat. To make it longer, ask "Continue the clip: now a surfer paddles out" or "Make the clip longer": the next part starts from the clip's last frame, and the two are joined into one clip. Each part adds up to 5 seconds. For a longer video, such as "a 1-minute video of a day at the beach", Toskar plans it as shots of up to 5 seconds and shows you the shots and about how long they take before anything is made. Each shot starts from the last frame of the one before, and they're joined into one video. A minute takes hours, so it's best left running.

What you need

  • An Admin account: only an Admin opens Tools and Profiles (People).
  • For a tool from an MCP server, sometimes a program such as Node.js or uv, or a sign-in to the service. Adding it says which.

See your tools

  1. Open Tools.
  2. Your tools lists every tool, built in or added. Search, or choose Built-in, Added, or Disabled to narrow the list.
  3. Choose a tool to see its details, then Disable to keep it out of every chat, whatever a profile allows.
The Tools page's Your tools tab: cards for Web Search, Open Web Page, Find Places, Create File, Analyze Spreadsheet, Git Commit, Run Code, List issues, and Search, each marked Enabled or Disabled
Each card says what the tool does and where it comes from: built in, a connected service, or a tool source.

Each message is offered only the tools it needs: web search for current questions, files when a message names a file, the terminal for "run" or "install". A plain question gets none, and the model can't call a tool it wasn't offered. How many tool calls a message may make depends on effort: 3 on Fast, 10 on Balanced, and 16 on Thorough.

Choose when a tool asks first

A profile decides which tools its chats may use (Profiles). Turning on a capability, such as Files or Run code, allows its tools. In advanced mode, each tool has its own permission:

PermissionWhat happens
Always allowIt runs without asking
Ask each timeChat asks before every run
Ask once per sessionChat asks the first time, then lets it run for the rest of the chat
DenyIt never runs

Tools that only read, such as web search, usually run without asking. Tools that change something, such as writing a file or making a Git commit, ask first, and you can choose Allow for session. Git push asks every time, since it sends your code off this computer.

Settings → Tool permissions sets a limit for the terminal, file writes, and Git on this computer. A profile can be stricter than Settings, never looser, so a profile set to Always allow still asks while Settings says Ask.

Answer when a tool asks

When a tool asks first, chat shows what it wants to do and why, and waits.

A dialog over a chat asking Allow Write file?, with the reason, the file name budget-by-month.csv, the details, and Allow once, Allow for session, and Deny buttons
The question shows what will change and why, before anything happens.
  1. Read what the tool will do, such as the file it will save or the command it will run.
  2. Choose Allow once to let it run this time, Allow for session to let that tool run for the rest of this chat, or Deny.

Some things ask every time, even after Allow for session: a command that can't be undone, such as deleting files, and any change right after the chat read a web page or a file from outside. What a page or file says is data, not instructions, so it can't approve anything on your behalf.

Tip: A chat in another app, such as the terminal or the desktop app, asks in whichever app you have open. The first answer counts, and the others put the question away.

Add tools from an MCP server

MCP is a common way for apps to offer tools to an AI. Toskar can use tools from MCP servers, such as ones for Notion, Linear, a database, or a folder.

  1. Open Tools, then Add more, then Add tools.
  2. Choose a way in. Gallery has ready-made servers. Paste takes whatever a server's instructions say. From your other apps reuses servers you set up elsewhere. Custom takes a command or a web address.
  3. Answer what it asks, such as a folder, a key, or a sign-in.
  4. Toskar connects and says what it found, such as "ready with 14 tools: 10 that read, and 4 that change things and ask you first".

Nothing is saved if the server can't be reached, and the message says what to do, such as installing Node.js. Keys and tokens are kept apart from Toskar's database and are never shown to the model. Tool sources, on the Add more tab, lists each server with its tools, its log, and a Sign in action for services that need one.

Use Toskar from other AI apps

Toskar is an MCP server too, so apps such as Claude Desktop, Cursor, and VS Code can ask your local AI and search your connected knowledge. API Access, under Use Toskar in other AI apps, shows what to paste into each. The local AI only uses tools that read when another app asks it. Here are the tools Toskar offers:

curl -s http://127.0.0.1:7331/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}'

You’ll see something like:

{…"name":"ask_local_ai"…"name":"list_local_models"…"name":"search_my_knowledge"…}

Limits

  • Run code uses the built-in sandbox on macOS, and needs bubblewrap on Linux. It isn't available on Windows yet.
  • A tool from an MCP server on the web reaches that service. Each call shows in Settings, under What left this computer.
  • Small models sometimes pick the wrong tool, or answer without one. A larger model does better.

Troubleshooting

What you seeWhat to do
A tool never runsCheck that it's on in Tools, and allowed in the chat's profile
Chat asks every time, even after Allow for sessionThe command can't be undone, or the chat just read something from outside. That always asks
An MCP server says it needs Node.js or uvInstall the program it names, then add the server again
A server's card says Sign in to use itChoose Sign in on its card, and finish the sign-in in the browser window
The answer didn't use a tool you expectedName the tool or its subject in your message, or raise effort for more tool calls

These pages describe Toskar 1.14.1. Download 1.14.1 and its release notes ↗ · Need help? Report a bug or request a feature ↗

Something missing or wrong?

Kind

Sent with it: docs 1.14.1. It becomes a public issue in toskar-core.

Tools, permissions, and connected services · Toskar documentation