OpenClaw · Cobble Docs

OpenClaw is an open-source personal AI assistant. Cobble plugs in as a custom provider — one config block, plus one gotcha (the model allowlist) that this guide covers so you don't hit it.

Works best with: Pro or Max. OpenClaw runs continuously, so parallel slots and a deep weekly budget matter more than raw model size.

1. Add Cobble as a provider

In your OpenClaw config, add Cobble under models.providers:

json

{
  "models": {
    "providers": {
      "cobble": {
        "baseUrl": "https://api.cobble.network/v1",
        "apiKey": "sk-cobble-...",
        "api": "openai-completions",
        "models": [
          {
            "id": "qwen/qwen3.8-27b",
            "name": "Qwen3.8 27B",
            "contextWindow": 128000
          }
        ]
      }
    }
  }
}

2. Allowlist the model — don't skip this

OpenClaw rejects any model not listed in agents.defaults.models, using the format provider-name/model-id:

json

{
  "agents": {
    "defaults": {
      "models": ["cobble/qwen/qwen3.8-27b"]
    }
  }
}

If you skip this step, the provider config looks correct but every request is refused — it's the most common setup failure.

3. Restart and select

Restart OpenClaw, then set your default model to cobble/qwen/qwen3.8-27b.

Troubleshooting

"Model not allowed" — step 2. The allowlist entry needs the cobble/ prefix in front of the full model ID.
400 errors from the API — make sure api is "openai-completions". OpenClaw automatically disables OpenAI-only request fields for non-OpenAI hosts, but only when the API type is right.
401 — key issue; test with curl against /v1/models first.
429 overnight — an always-on assistant with heartbeats can drain a 5-hour budget while you sleep. Check the window field in the 429 body: 5h resets within hours, weekly means the week's budget is spent. Either lengthen OpenClaw's polling interval or set a small wallet spend cap as a buffer.

Plan notes

OpenClaw is the most usage-hungry tool we support — it's an assistant that never logs off. Heartbeats with a warm prompt cache cost a fraction of a cent each, so a Pro budget comfortably covers normal conversational use, but aggressive heartbeat/cron configurations against a large uncached prompt are the classic way to hit 429s. Tune the intervals before upgrading plans.