Skip to content

Connect a model provider

Nerve gets its current built-in provider and model catalog through its pi-ai integration. The Workbench shows the providers and authentication methods available in the installed version; the underlying library’s complete transport support is not automatically a Nerve product promise.

Open Settings → Providers and choose a provider. Depending on provider metadata, Nerve can offer:

  • an API key;
  • an OAuth or subscription flow;
  • both methods as alternatives.

A provider has one active credential type. Saving an API key replaces OAuth for that provider, and completing OAuth replaces its API key.

OAuth-capable options vary by release and provider. Current code exercises flows for Anthropic, GitHub Copilot, Kimi Coding, OpenAI Codex, OpenRouter, Radius, and xAI, but consult the installed provider dialog rather than treating this as a permanent support matrix.

After authentication, choose:

  • a global default model and thinking level;
  • an optional scoped list for the conversation picker;
  • a separate default model and thinking level for Explore sub-agents.

An empty scope means all authenticated models. If a scoped model later disappears, Nerve retains it as unavailable but ignores it in the picker.

Thinking choices are model-dependent. Nerve’s complete level vocabulary is off, minimal, low, medium, high, xhigh, and max; the runtime clamps a choice to levels the selected model reports supporting.

The composer can change model and thinking level without recreating a conversation. If a run is active, the change applies to the next model request, not the request already streaming.

Some providers expose only the first eight models in Nerve’s catalog. Full lists are currently enabled for a limited set of provider families. Add a custom model definition when a compatible model is not listed.

Voice transcription requires OpenAI Codex/ChatGPT OAuth. An ordinary OpenAI API key is not enough. Audio is uploaded to ChatGPT’s transcription service.