@brglng/pi-model-discovery

A Pi extension that adds model auto-discovery to Pi using only the providers section of Pi's native ~/.pi/agent/models.json — no extension development needed. Same format as Pi's built-in model configuration, with <baseUrl>/models discovery and per-model overrides built in.

Installation

pi install /absolute/path/to/pi-model-discovery

After publication:

pi install npm:@brglng/pi-model-discovery

Configuration

Providers are defined under the providers key of Pi's native ~/.pi/agent/models.json. The key is the provider ID; each value is one provider. For example:

{
  "providers": {
    "gateway": {
      "$schema": "https://raw.githubusercontent.com/brglng/pi-model-discovery/main/schemas/config.schema.json",
      "name": "My Gateway",
      "baseUrl": "https://api.example.com/v1",
      "apiKey": "$GATEWAY_API_KEY",
      "api": "openai-completions",
      "models": [
        {
          "id": "special-model",
          "api": "openai-responses",
          "baseUrl": "http://192.168.5.46:18000/v1",
          "maxTokens": 32768
        }
      ]
    }
  }
}

baseUrl is the provider default URL and is optional. Providers without a baseUrl remain in the configuration, but the extension skips remote model discovery for them. apiKey uses Pi's native config value syntax. API keys are provider-level; Pi's native model configuration does not support a separate API key per model. It supports literals, $ENV_VAR, ${ENV_VAR}, and leading !command values. The key is never stored in the configuration when an environment variable or command reference is used.

Model discovery and overrides

Model discovery is always enabled: the extension requests <baseUrl>/models with a Bearer token. It accepts common OpenAI-style responses (data), bare arrays, and catalogs nested under models or output.models. Set PI_OFFLINE=1 to skip all network requests during startup.

Server model metadata is used by default. Entries in the models array can specify overrides for configured model IDs. Provider and model fields are kept as supplied; this extension does not validate fields it does not consume. Configured model IDs are also retained when the server does not return them. Common fields include name, api, baseUrl, reasoning, thinkingLevelMap, input, cost, contextWindow, maxTokens, samplingParams, headers, and compat. This allows each model to select a different endpoint and API type such as openai-completions, openai-responses, or anthropic-messages. The spelling anthropic-message is accepted as an alias.

If discovery fails, explicitly configured models remain available.

Commands

• /model-discovery status shows configured providers and model counts.
• /model-discovery refresh refreshes every catalog.
• /model-discovery refresh <provider> refreshes one catalog.

License

MPL-2.0