Models

Uber uses pi’s built-in providers, with API keys from the environment (OPENAI_API_KEY, ANTHROPIC_API_KEY and so on). Choose the default model and the models a session can switch to in settings.json, as in pi:

{
  "defaultProvider": "openai",
  "defaultModel": "gpt-6-sol",
  "enabledModels": ["openai/*", "anthropic/*"]
}

Add other providers, such as an OpenAI-compatible gateway, in models.json, as in pi:

{
  "providers": {
    "gateway": {
      "baseUrl": "https://llm.example.com/v1",
      "api": "openai-completions",
      "apiKey": "$GATEWAY_API_KEY",
      "models": [{ "id": "claude-sonnet-5-5", "reasoning": true }]
    }
  }
}

Providers that extensions register with pi.registerProvider() work the same way. A turn’s model switches the session’s model, as provider/id. The session keeps it for later turns, and subagents use it:

rebuno exec create uber '{"task": "Review the change.", "model": "openai/gpt-6-sol"}' --session fix-tests

A turn’s thinking sets the session’s thinking level the same way: off, minimal, low, medium, high, xhigh, or max. A model that reasons less uses its nearest level. Mark a models.json model that reasons with "reasoning": true; any other model uses off.

Recording

Every model call is a recorded llm_call whose target is the model ID, so policy can allow, deny, or require approval per model. Providers work when their API accepts a custom fetch: openai-completions, openai-responses, azure-openai-responses, anthropic-messages, and mistral-conversations. Models on other APIs, such as Google’s, Bedrock, and Codex, are left out.

Get started

Configure

Operations

Recording

​Recording

Recording