AI providers

co-maintainer asks an AI provider to write the guides in init and sync and to write every review. Five providers are supported.

--ai=none is the default setting until you choose a provider. It turns AI off, which is enough for probe because it needs no model. init, sync and review all refuse it: the guides and the reviews are written by the provider, so there is nothing to build without one.

Where each provider works

Every provider works everywhere. A review with no provider saved uses OpenRouter, which is what every review used before other providers existed, and review --ai=none is refused. For a remote review the provider is whichever one the server is configured with, so the laptop never needs a key unless it brings its own (see your own key).

Pick a provider and save it once

co-maintainer set --ai=anthropic --token=YOUR_KEY \
  --low-model=YOUR_LOW_MODEL_ID --high-model=YOUR_HIGH_MODEL_ID

After that init, sync and serve need no flags. The dashboard does the same under Settings, Models and API key.

There is no default model

co-maintainer never picks a model for you. Models change quickly, and an installed version cannot fetch a newer list, so a built-in default would one day name a model the provider no longer serves. You choose both models once and save them, or pass them on the command:

When one is missing, the command stops before it spends anything:

• In a terminal it asks for the value and suggests none.
• Without a terminal, or with --json, it exits with code 2 and says which one is missing: Missing low model. Pass it as a CLI option when running without an interactive terminal.
• A review that reaches a model it does not have reports No low model is configured. with the hint Run co-maintainer set --low-model=..., or pass --low-model=... on the command that needs it.

Use the model ids the provider lists, for example the ones on its models page. The examples in these docs use an OpenRouter pair such as openai/gpt-oss-120b and openai/gpt-5.6-luna. They are examples, so check your provider's current list before you copy one.

Keys

OPENROUTER_API_KEY is only read when the provider is OpenRouter and OPENCODE_API_KEY only for OpenCode, so a key for one provider is never sent to another. The saved key belongs to the saved provider, and every command, review included, sends it there. OPENROUTER_LOW_MODEL and OPENROUTER_HIGH_MODEL are likewise only read for OpenRouter.

serve reads the same variables for its jobs, so a server can keep its key in the environment instead of config.json. The Settings page then says the key comes from the server's environment. A key is never stored per repository. The order every setting is resolved in is on Configuration.

set checks the key and the model

Before it writes anything, set asks the provider whether the key works and, if you gave a model, whether the provider has it:

A rejected key or an unknown model exits with code 2 and nothing is saved. A network failure only warns, because that is not a typo. Pass --no-verify to skip the check.

What differs between providers

• OpenRouter reports the dollar cost of every call, so its reviews show a real cost.

• OpenAI is called through its Responses API. It does not report a cost, so the cost of a review is unknown. During a codegraph tool conversation the reasoning items of a reasoning model are not passed back between rounds, which can make a later round reason a little less well. The number of rounds is small, so little is at stake.

• Anthropic is called through its Messages API. It does not report a cost either. The guides are the same prefix in every review and are marked for prompt caching, so a review that follows soon after another reads them from Anthropic's cache at the cached rate. The call is not streamed, so a very long answer can run into Anthropic's request timeout.

• OpenCode Zen and Go route each model to its own API format, and the model list does not say which. co-maintainer calls every model over Chat Completions. A model that only answers in another format is refused with OpenCode Zen does not know the model and a hint that says so. Pick a model that supports Chat Completions. Gemini models on Zen use Google's format and do not work. Neither gateway reports a cost, so a review on it is unknown.

• OpenCode Go is a subscription meant for coding agents, and OpenCode watches its traffic for abuse. co-maintainer identifies itself with its own user agent and one session id per run, as Go asks every client to. init makes many short calls in a row, which is not the traffic Go is built for, so heavy use may hit its limits or its abuse checks. Zen has no such rule.

unknown is not zero. Cost explains how the two are kept apart. The probe estimate says estimate unavailable for this provider for OpenAI and Anthropic instead of printing a made-up dollar range.

Experimental: Sign in with ChatGPT

review, init and sync can run on a ChatGPT subscription instead of an API key, through OpenAI's Sign in with ChatGPT. Add --experimental-siwc to the command. --ai and --token are then unused, and neither can be combined with it.

Sign in once, then check it:

co-maintainer oauth login --provider=sign-in-with-chatgpt
co-maintainer oauth list-models --provider=sign-in-with-chatgpt

--provider=siwc is accepted as a short alias. oauth logout removes the saved credentials (useful before switching accounts), oauth refresh renews the token, and oauth ping prints the account and its models. Credentials live in ~/.config/co-maintainer/oauth/sign-in-with-chatgpt.json, readable only by you.

At least a ChatGPT Go, Plus or Pro plan is required. A Free account is not eligible, and buying credits or signing in again does not change that. See the OpenAI help article Using your ChatGPT plan in other apps and sites.

Models still come from --low-model and --high-model. The account's own list is printed when one is missing or unknown, and a leading openai/ in a model id is ignored, so an OpenRouter id like openai/gpt-5.6-luna also works here.

This is experimental. It is CLI-only for now, because the sign-in needs a browser: serve, the dashboard and remote review do not use it. A ChatGPT call reports no dollar cost, so a review on it shows unknown.

Errors

More in Troubleshooting.