Does OpenClaw (Lobster) native model work?

OpenClaw can connect to local models, but don't assume that "being able to chat" is the same as "being able to be an agent". If a small request works on the local OpenAI-compatible backend but OpenClaw agent turn fails, it is usually due to an incompatibility in the message format, tool call, or context length.

Do a three-layer test first

1. Directly hit the local /v1/chat/completions to confirm that the backend is available.
2. Use openclaw infer model run to test whether the model can complete the basic reply in OpenClaw.
3. Let the agent call up a simple tool, such as reading a small file or executing a security command.

Common compatible amendments

• Report messages[].content String required: Add compat.requiresStringContent: true to the model.
• Normal chat works, tool calls fail: Consider setting up compat.supportsTools: false and don't let it take on tool-based tasks.
• Small prompts can crash, large contexts crash: This is a model or server-side limitation, reducing the context or changing the backend.

Local models are suitable for privacy, offline, and low-cost scenarios, but complex toolchains are more stable for the model. It is recommended to put the local model in low-risk tasks or fallbacks, and the key automation is still covered by a strong model.

Official open source address: https://github.com/openclaw/openclaw.

The safest approach is to build a separate test agent for the local model, allowing only low-risk tools. Make sure it can handle contexts, tool calls, and error recovery stably, and then put it in the main agent's fallback.