TarsTARS

Models and Sign-in

Sign in to any pi registry provider and pick a discovered model and thinking level.

Tars runs on the pi model registry, so every provider pi knows about is available: OpenAI Codex (ChatGPT subscriptions), Anthropic, OpenAI, Google, GitHub Copilot, OpenRouter, and many more, plus local and custom OpenAI-compatible endpoints.

Pick a provider and model

tars model

The command walks through four steps and writes the result to ~/.tars/config.json:

  1. Provider — the list shows each provider’s sign-in methods (OAuth, API key, or environment credentials) and whether you are already signed in.
  2. Credentials — keep the current credentials, sign in with OAuth in your browser, enter an API key, or import an existing login from the standalone pi CLI (~/.pi/agent/auth.json).
  3. Model — models are discovered from the signed-in account (the live catalog is refreshed when the network allows). You can always enter a model id manually.
  4. Thinking level — reasoning-capable models ask for off, minimal, low, medium, high, xhigh, or max; only levels the model supports are offered and medium is suggested. The context window is taken from the catalog.

Apply the change with tars restart.

Non-interactive use

Flags pre-answer prompts, which is handy over SSH or in scripts. With --provider, stored credentials for that provider are kept as they are (re-authenticate with tars auth login), so the command finishes without prompting once you are signed in.

tars model --provider openai-codex --model gpt-5.6-luna --thinking medium
tars restart

--model requires --provider. An unsupported --thinking level for the chosen model fails before anything is written.

Sign in separately

tars auth login openai-codex        # OAuth (ChatGPT Plus/Pro)
tars auth login anthropic --api-key # API key instead of OAuth for providers that offer both
tars auth status
tars auth logout openai-codex

Credentials live in ~/.tars/auth.json with owner-only permissions. OAuth tokens refresh automatically; the supervisor picks up new credentials on the next message without a restart. Providers without an interactive login (for example cloud SDK profiles) read ambient credentials from the environment.

Configuration keys

JSON Environment Purpose
piProvider PI_PROVIDER Provider id, for example openai-codex
piModel PI_MODEL Model id, for example gpt-5.6-luna
piThinkingLevel PI_THINKING_LEVEL Reasoning effort; ignored by non-reasoning models
piBaseUrl PI_BASE_URL Local or custom OpenAI-compatible endpoint

Local and custom endpoints keep their optional key in ~/.tars/.env (LOCAL_API_KEY or CUSTOM_API_KEY). When a thinking level other than off is set, Tars advertises reasoning for the endpoint model so pi forwards the effort parameter.