Choose a Local Model for OpenClaw

A local model is useful when prompts and inference must stay on hardware you control. It is not automatically the best agent model: OpenClaw also needs reliable structured tool calls, sufficient context, and recovery after a failed action.

Local runtime requirements

For Ollama, confirm the server and provider catalog before changing OpenClaw:

ollama list
openclaw models list --provider ollama

Use the native Ollama base URL without /v1. OpenClaw’s native integration preserves tool-call behavior that can be lost through an OpenAI-compatible compatibility endpoint.

Match the model to the machine

Leave memory for the operating system, Gateway, context cache, tools, and any desktop applications. A model that barely loads can become unreliable once a long conversation or browser task expands the working set.

Use the dedicated 16 GB and 32 GB guides as starting budgets, then measure your own machine.

Require native tool support

A model that answers chat prompts well can still emit malformed tool arguments or describe a tool instead of calling it. Run the tool-calling test before making it primary.

Check the real privacy boundary

Local inference keeps model requests on the configured runtime, but Skills and tools may still call web services, APIs, chat channels, or remote browsers. Review the whole agent path before calling a setup offline or private.

Set and verify

openclaw models set <ollama/model-id>
openclaw models status

Expected result

The selected model fits without sustained swapping, completes native tool calls, and remains responsive through the longest representative task you plan to run.