OpenAI-Compatible Models with OpenClaw

“OpenAI-compatible” describes an API shape, not guaranteed OpenClaw agent compatibility. A server may pass a small chat-completions request and still fail on tool schemas, structured content, long prompts, or streaming.

Verify the endpoint first

Confirm the base URL, authentication method, model identifier, and API path from the server’s own documentation. Test one small direct request before adding OpenClaw. Keep credentials outside committed configuration.

Then inspect discovery and status:

openclaw models list
openclaw models status
openclaw logs --follow

Separate common compatibility failures

• A direct request fails: fix the endpoint, server, credential, or model name.
• Direct text succeeds but OpenClaw rejects message content: review provider compatibility settings such as string-only content.
• Text succeeds but tools fail: verify native tool support and the provider’s tool-schema compatibility.
• Small prompts work but agent turns fail: check real context and server memory limits.
• Raw tool JSON appears as text: use the provider’s native OpenClaw integration when one exists.

Do not mark unsupported tool calling as supported simply to pass discovery. That can make an agent appear healthy while it silently answers instead of acting.

Run a bounded qualification

Use the same tasks from Test model tool calling: read, write, harmless command, multi-step transformation, and invalid-input recovery.

Expected result

• The exact model is discoverable.
• A fresh OpenClaw session completes a normal turn.
• Tool calls contain valid names and arguments.
• The endpoint handles the configured context without crashing or truncating required instructions.

Record the server version and compatibility overrides with every published result.