| name | openclaw-agents |
| description | Use this project skill when Michael asks to configure OpenClaw providers, add a local LLM, switch model, configure Qwen3.5, inspect OpenClaw tools, subagents, system prompts, Pi SDK agent runtime, or model provider config.
|
OpenClaw Agents And Providers
Scope
Use this skill for OpenClaw/OpenIntellisoins agent configuration from this
machine. The main OpenClaw repo is expected at ~/openclaw/openclaw/, with
runtime config at ~/.openclaw/openclaw.json.
Load references/providers-catalog.md only when the user asks about provider
types, discovery, failover, or the provider registry files.
Current Local Provider
Primary local provider:
{
"providers": {
"mlx-qwen35": {
"type": "openai-compatible",
"baseUrl": "http://localhost:8087/v1",
"model": "qwen3.5-35b-a3b",
"contextWindow": 262144,
"reasoning": true
}
},
"agents": {
"default": { "provider": "mlx-qwen35" }
}
}
Check the local model server before debugging OpenClaw agent behavior:
curl http://localhost:8087/v1/models
Key Files
| File | Role |
|---|
src/agents/pi-embedded-runner/run.ts | Agent runtime lifecycle |
src/agents/tool-catalog.ts | Tool catalog |
src/agents/tool-policy.ts | Tool allowlists |
src/agents/subagent-registry.ts | Subagent registry |
src/agents/subagent-spawn.ts | Subagent spawning |
src/agents/system-prompt.ts | Dynamic system prompt |
Troubleshooting
For Qwen3.5 failures:
- Check
curl http://localhost:8087/v1/models.
- If the call times out, the model may still be loading.
- If OpenClaw hangs, inspect gateway logs and active connections to port
8087.
- If tool calling fails, verify provider
reasoning: true and OpenAI-compatible
tool call support.
