Configure your model

An llm cell reaches its provider through three params: model, api_key and base_url. provider names the wire protocol and not the vendor, so "openai" means the OpenAI-compatible HTTP API and base_url decides who answers it. The shipped templates write those params as ${VAR} tokens and read the values from the colony’s .env.

A ${VAR} binds at every read, not at instantiation. The token stays literal in the tree, the secret stays in one file, and another model is an .env line and a restart, not a config edit.

Set it

.env sits next to the colony root; --env <path> points at another file.

OPENROUTER_API_KEY=sk-...
MODEL_BRAIN=openai/gpt-5.6-luna

OPENROUTER_API_KEY binds api_key in templates/talky/brain/config.json. It has no default, so the grow step needs the variable to exist: a missing one is a hard reject, env_var_missing, naming it. MODEL_BRAIN is read by examples/meclaw-os/grow.json as "ctx": {"model": "${MODEL_BRAIN}"} and arrives in the brain as ${ctx.model}. The assistant has two brains, and grow-cogny.json passes MODEL_CORE the same way, so the surface can run a fast model while the core runs a slow one (why/two-brains.md).

Another endpoint

Any endpoint on the same wire works, and OpenRouter is only the default base_url the templates carry. Some examples declare it as a token (OPENROUTER_BASE_URL in examples/telegram-research); talky and cogny name the endpoint outright, so the manifest that grows the cell overrides it. An empty api_key is no credential, and the cell then sends no Authorization header at all, which is what a local server on loopback wants.

{"name": "talky", "template": "talky",
 "override_params": {"brain": {"base_url": "http://127.0.0.1:11434/v1", "api_key": ""}}}

Read on

cell-types.md has every llm param and config.md the substitution rules. costs.md reads what a colony spent per model out of colony.db.