Explained
What's an agent harness, and how Toskar's works
A model on its own only writes the next words. The agent harness is the code around it that makes it an agent: it gives the model tools, runs the calls it makes, decides what it may do, and keeps track of the work. Toskar Core includes one for the models on your own computers, and lets you bring your own.
The parts
What Toskar's harness does on each turn
The same harness answers in chat, on your phone, in the terminal, and through the API.
The loop
Toskar sends the model the conversation and the tools it may use, runs the tool calls it makes, hands back the results, and stops when there's an answer or a limit is reached: 3 tool calls on Fast, 10 on Balanced, and 16 on Thorough.
Only the tools a message needs
Web search for a current question, files for a folder, the terminal for "run" or "install", Git for "commit". A plain question gets no tools, and a model can't call a tool it wasn't offered. MCP servers add more.
Permissions
Each tool is Always allow, Ask each time, Ask once per session, or Deny. After an answer reads a web page or a file, a tool that changes something asks first, and text that gives orders to an AI is left out before the model reads it.
Memory and knowledge
What you asked it to remember, and the passages from your own files and databases that match the question, are added to what the model reads, with sources it can cite.
Checks and retries
A profile can check an answer and report or correct it, try other models in order when one fails, and retry a failed turn on another computer with the same model.
Context
Long chats are compacted automatically, and a profile sets how much of the model's window earlier messages may use, so the request still fits.
Orchestration
One model, or a team of them
A profile is a saved assistant: its models, tools, memory, and the computers it runs on. Its strategy decides how a request is worked through.
Auto
Quick questions get one model. Bigger requests are planned and checked.
Single model
One model answers everything, with the profile's tools.
Planner + workers
A request with several parts is always worked through in parts.
Team
A planner splits the request, a worker does each part, and a reviewer checks the answer. With paired computers, the parts run on different machines at the same time.
The built-in Programming profile uses Team. Each role can have its own model and computer, and in advanced mode Run details under an answer shows which model ran each part, where, and how long it took.
Bring your own
Or use your own harness on Toskar's models
Coding agents such as Claude Code, Codex, and OpenCode are harnesses too. Point one at Toskar and it drives a model on your computer directly, with its own tools, through the OpenAI-, Anthropic-, or Responses-compatible API. Local AI for developers has the details.
In a terminal, toskarctl chat uses Toskar's own harness: asked to do something in a folder, such as fixing a failing test, it works step by step and asks before it changes anything.
What to get
The products that fit
Questions
- What's the difference between a harness and an orchestrator?
- People use the words loosely. Here, the harness is everything around one model's turn: the loop, the tools, permissions, memory, and context. Orchestration is how a request is split between models and computers: who plans, who works on each part, and who checks. Toskar does both, and a profile's Strategy and Orchestration settings choose how.
- Can I use Claude Code, Codex, or OpenCode instead of Toskar's harness?
- Yes. Name a model as raw: and its ID, and the request goes straight to the model with your agent's tools, and the model's tool calls come back to the agent. Claude Code uses the Anthropic-compatible /v1/messages, Codex the Responses-compatible /v1/responses, and toskarctl connect prints or writes the settings for Claude Code, Codex, OpenCode, aider, and several editors.
- Does it work with local models' tool calling?
- Yes, with any installed model that supports tools. How well an agent does depends on the model: small models often lose track over many steps, and larger ones hold up better. An agent with many tools still fits: when its tools would take more than a quarter of what the model reads at once, the model gets the ones the request most likely needs and searches for the rest.
- Does anything leave my computer?
- Models run on your own computers. Web search, the browser, and connected services reach the internet only when a profile allows them, and Settings lists everything that left, under What left this computer.
- Is it open source?
- Yes. Toskar Core, harness included, is free software under AGPL-3.0-or-later, and the source is on GitHub.