OpenAI-compatible local APIs with PicoClaw

Many local and self-hosted stacks expose an OpenAI-compatible HTTP API (same JSON shape, different base URL). If PicoClaw supports your chosen provider or custom base URL in config, you can route assistants through LM Studio, vLLM, LiteLLM, or an internal gateway—without changing your automation code.

1. Common setups

  • Desktop or workstation: LM Studio or Ollama with a compatible binding; PicoClaw on the same machine or LAN.
  • Homelab GPU node: vLLM or Text Generation Inference behind nginx; PicoClaw on a small VM or Pi calling the LAN URL.
  • Unified proxy: LiteLLM in front of several backends; one API key and model alias for PicoClaw.

2. Networking and TLS

Use HTTPS on anything exposed beyond localhost. If PicoClaw runs in Docker, prefer the service name as hostname instead of localhost from inside the container. Firewalls should allow only the subnets that need to reach the model server.

3. Configure PicoClaw

Set the provider and api_base (or equivalent) to your local endpoint, then pick the model id your server advertises. Cross-check field names on Configuration and examples in Providers. For Ollama-native endpoints, you may prefer the dedicated Ollama guide.

4. Performance on small hardware

If the model host is a Raspberry Pi, stick to small quantised models and short contexts. Often the best pattern is: Pi runs PicoClaw; a PC or NAS runs the LLM.

5. Next steps