Skip to content

Models

The agent can use several models. You describe them in the models section and say which one does what.

Providers

A provider is where a model comes from. There are two types.

  • claude-cli: an unmodified claude command-line tool, installed and logged in on the machine, using your own subscription. No key goes in the configuration. An optional command names the binary.
  • openai: any OpenAI-compatible Chat Completions service: OpenAI, OpenRouter, a local Ollama. It needs base_url, and key unless the address is local (localhost, 127.0.0.1).
models:
  providers:
    - {name: claude, type: claude-cli}
    - {name: openrouter, type: openai, base_url: https://openrouter.ai/api/v1, key: env:OPENROUTER_API_KEY}

Models

Each model entry says what the model is:

  models:
    - {id: sonnet, provider: claude, name: sonnet, tools: true, context: 200000}
    - {id: flash, provider: openrouter, name: deepseek/deepseek-v4.1-flash, tools: true, context: 1048576, price_in: 0.045, price_out: 0.5}
  • id is the short name that purposes use. It defaults to name.
  • name is the provider's own name for the model.
  • tools says the model can call tools. vision says it reads images. local says the data stays on your hardware.
  • context is the window in tokens.
  • price_in and price_out are US dollars per million tokens. Use 0 for subscription and local models.

Purposes and fallback

A purpose is a job. Each lists its models in the order they are tried. If a model is out of rate limits or fails, the next one is used.

Purpose What it does If not set
chat A conversation turn. Required.
check Quick yes/no checks in a clean context. Uses chat's list.
compact Summaries of long history. Uses chat's list.
translate The bot's fixed texts in a new language. Uses chat's list.
memory The memory review. Uses check's list.
code Writing modules and skills. No fallback.
  purposes:
    chat:    [{model: sonnet, effort: medium}, {model: haiku}]
    check:   [{model: haiku, effort: low}]
    code:    [{model: opus, effort: high}, {model: sonnet, effort: high}]

effort is low, medium or high, and also xhigh or max on the claude-cli route.

Logging in the claude CLI

On your own machine, log in with claude as you normally would, as the user that runs the node. ganglion never sees the login.

On a server or in a container, where you cannot log in interactively, run this on a machine where you can:

claude setup-token

It prints a token. Set it as CLAUDE_CODE_OAUTH_TOKEN in the node's environment. The claude-cli provider passes it to the CLI. Commands the agent runs do not get it.

A login through the chat is not built yet. Use your own subscription for your own node only. Taking a subscription token and calling the API with it directly is not supported.

Limits

Spending is bounded per day, in the node's time zone (node.tz).

  • agent.limits.daily_turns (default 300) and daily_tokens (default 1,000,000) apply per person. Past either, the bot answers "Daily limit reached; the bot answers again tomorrow." and alerts the admin. Scheduled tasks count too.
  • agent.limits.daily_budget_usd (default 1) caps spend. Past it, a turn uses only free and local models (price 0). Background housekeeping calls are not cut. Models on a subscription cost 0, so the budget does not limit them unless you set price_in and price_out.

/status shows today's turns and tokens against the limits and what was spent. Set a limit to 0 or off to remove it.