Skip to content

OpenClaw Integration

Jump to bottom
youngharold edited this page Feb 18, 2026 · 4 revisions

OpenClaw Integration

Once the Tightwad cluster or speculative proxy is running, register it as a provider in OpenClaw so agents can use the models.

Provider Registration

Add to ~/.openclaw/openclaw.json under providers:

"tightwad": {
  "baseUrl": "http://<rocm-machine-ip>:8080/v1",
  "apiKey": "no-key",
  "api": "openai-completions"
}

For the speculative decoding proxy:

"tightwad-proxy": {
  "baseUrl": "http://127.0.0.1:8088/v1",
  "apiKey": "no-key",
  "api": "openai-completions"
}

Model Registration

Add models under models (or agents.defaults.models):

"tightwad/qwen3-32b": {
  "id": "qwen3-32b",
  "name": "Qwen3 32B (Tightwad Cluster)",
  "provider": "tightwad"
}

For the proxy (output quality = target model):

"tightwad-proxy/qwen3-32b": {
  "id": "qwen3-32b",
  "name": "Qwen3 32B (Speculative Proxy)",
  "provider": "tightwad-proxy"
}

Assigning to Agents

Override specific cron jobs or agents to use Tightwad models:

{
  "model": "tightwad/qwen3-32b"
}

Since Tightwad serves an OpenAI-compatible API, it works the same as any other provider in OpenClaw. No special configuration needed beyond the provider/model registration.

Coexistence with Existing Providers

Tightwad runs alongside existing providers:

  • llama-desktop (GLM-4.7-Flash on port 8080 of Windows desktop) — may need to change ports if both run simultaneously
  • ollama providers — unaffected, different ports
  • anthropic / anthropic-fleet — cloud providers, no conflict

Port conflict note: The Windows desktop currently runs llama-server on port 8080 for GLM. If Tightwad's coordinator also uses 8080 (on the ROCm machine), there's no conflict since they're different machines. But the Windows rpc-server instances share the desktop's GPUs — running both GLM llama-server and rpc-server simultaneously may cause GPU memory pressure.

Model Hot-Swap via OpenClaw

If you swap models with tightwad swap deepseek-r1-70b, the API endpoint stays the same but serves a different model. OpenClaw agents calling tightwad/qwen3-32b would get DeepSeek responses. To handle this cleanly, register each model as a separate OpenClaw model and only start the one you need, or implement a routing layer.

Clone this wiki locally