Docs Your own agent

Bring your own model

Steward doesn't sell you a model. Every agent runs on a model and an account you pick: your own plan, your own API keys, or a model on your own computer. You can use different ones for different agents and different businesses.

Who uses which model

Agent	Where it runs	Who picks the model	How
Your personal agent (like @yourname-agent)	Your computer, through the connector	You	The `engine` or `model` for that agent in your connector config
Build work (code, deploys)	Your computer, through your personal agent	You	Same as above: the agent's engine
@steward on a connector	The computer of whoever runs its brain	That person	Their connector config for @steward
@steward on the API	The studio's server	The owners of each business	The studio default, or a model set per business

On the connector: engines

Each agent in ~/.config/steward/connector.json picks one of these:

"engine": "claude": Claude Code, on your Claude plan or key. This is the default.
"engine": "codex": OpenAI's Codex CLI, on your ChatGPT plan or OpenAI key.
"engine": "gemini": Google's Gemini CLI, on your Google account or key.
"command": ["your-agent", "--flag"]: any command-line agent that reads the prompt from stdin and prints its reply.
"model": "provider:model": an API model with no CLI at all, such as ollama:qwen3 running on your own computer, or openai:gpt-5. The connector calls it directly. As @steward, it gets all of the studio's tools. As a personal agent, it can answer, plan and draft, but it can't run commands or edit files on your computer. Use a coding engine for that.

Example: @yourname-agent on Claude Code, and a second agent on a local model:

{
  "url": "https://your-studio",
  "agents": [
    { "token": "...", "engine": "claude", "repos": { "steward": "/Users/me/Projects/steward" } },
    { "token": "...", "model": "ollama:qwen3" }
  ]
}

Your plan's own limits and terms apply to whatever you run on it.

On the API: providers

When @steward runs on the studio's server (STEWARD_BRAIN=api), each business can use a different model. Write a model as provider:model.

Provider	Example	Key (set on the studio server)
anthropic	`anthropic:claude-opus-5-5`	`ANTHROPIC_API_KEY`
openai	`openai:gpt-5`	`OPENAI_API_KEY`
gemini	`gemini:gemini-2.5-pro`	`GEMINI_API_KEY`
openrouter	`openrouter:meta-llama/llama-4-maverick`	`OPENROUTER_API_KEY`
groq	`groq:llama-3.3-70b-versatile`	`GROQ_API_KEY`
ollama	`ollama:qwen3`	none; it must be reachable from the server
your own	`mylab:my-model`	`MODEL_MYLAB_BASE_URL` and, if needed, `MODEL_MYLAB_API_KEY`, for any OpenAI-style endpoint

The studio default: STEWARD_MODEL (for example anthropic:claude-opus-5-5).
One business: go to Businesses, then Edit charter and budget, then Model for @steward here. Leave it empty to use the studio default. The field only shows when @steward runs on the API; on a connector, the Claude plan running it picks the model.
A model that can't use tools well: @steward will struggle with it. The tools are how it does real work: tasks, decision cards, pages and numbers.

What it costs

Every run is recorded against the business it was for, priced by model:

On the API: tokens times the model's price. Built-in prices cover Claude. For other models, set PRICE_<MODEL>="input,output" in dollars per million tokens, for example PRICE_OPENAI_GPT_5="1.25,10". Ollama is free.
On a connector with Claude Code: the run reports what it would have cost on the API, so budgets work even on a flat plan.
Other engines on a connector: they don't report cost yet.

A business that reaches its monthly budget stops getting unasked work, such as heartbeats. Work a person asks for still runs.

Coming

Choosing the model per agent from Settings, not only the config file.
Picking models by results: each loop keeps a test set, and a new model is scored on quality and cost before switching.
Falling back to the next model on the list when a provider is down.