Providers and login

Run radian login to store a key. Pick a built in provider from the numbered list. Pass custom to name a new endpoint. The key is never echoed. Radian verifies the key with a read only request to <base>/models first. Keys live in ~/.radian/credentials/auth.json. Run radian logout <provider> to remove one key. Removal asks first. Every other entry stays untouched.

Choose the provider with --provider: a built-in id (opencode-go, chatgpt, anthropic, openai, openai-compatible) or a [providers.<name>] name. An unknown provider is refused. Omit it to use openai-compatible. Choose the endpoint with --base-url. Omit --base-url to use the provider public endpoint. A custom --base-url or --header is refused for every non-generic provider. Choose the model with --model <name>. Use a bare model name. Never use <provider>/<model> on the command line. Use /model <provider>/<model> inside the full screen interface. Set model.provider, model.base_url, and model.model in the config file for defaults.

Pass --max-tokens <n> for Anthropic requests. The default is 16384. This flag is refused for openai-compatible requests. Pass --thinking <level> to set the thinking level. Use one of off, minimal, low, medium, high, xhigh, max. Omit it to use medium. Radian clamps the level to what the route supports. The older adaptive and enabled:<budget> spellings still work with a warning. On openai-compatible routes the level becomes the reasoning_effort request value. Use /thinking <level> inside the full screen interface to set the level.

Pass --header NAME=VALUE to add one request header. Repeat the flag to add more headers. Radian refuses Authorization here. The key comes only from the credential store. A value may hold {session_id}. Radian expands it to the current session id.

Radian looks up the context window from the models catalog. The catalog URL is read only. The cache lasts 24 hours. The fetch times out after 3 seconds. Pass --context-window <TOKENS> to set an explicit value. An explicit value always wins. Pass --offline-catalog to use only the local cache. Pass --catalog-provider <id> to look the model up under another catalog name. Omit it to derive the name from the base URL host.

Define named endpoints under [providers.<name>]. Set kind and base_url for each name. Set model.provider to the name to use it.

Read cli for every flag and exit code. Read configuration for every file key. Read environment for the key variables and lookup order.