Skip to content

OpenClaw Integration

OpenClaw is an open-source AI agent framework for macOS that provides tool-use, memory, and multi-step reasoning on top of a local OpenAI-compatible endpoint.

Squish exposes a fully OpenAI-compatible API at http://localhost:11435/v1, so OpenClaw can use any Squish-hosted model with zero code changes.

Quick Start

1. Start Squish

squish serve qwen2.5:7b

The server starts on http://localhost:11435 by default. It exposes: - POST /v1/chat/completions - GET /v1/models - POST /v1/completions

2. Configure OpenClaw

In your OpenClaw config file (usually ~/.openclaw/config.json or the project-level openclaw.json), point the agent at the Squish server:

{
  "agent": {
    "model": "openai/qwen2.5:7b",
    "openaiBaseUrl": "http://localhost:11435/v1",
    "apiKey": "squish"
  }
}

Note: Squish does not require an API key by default. Set "apiKey" to any non-empty string; the value is ignored unless you started Squish with --api-key.

3. Run OpenClaw

openclaw run "Summarise the README in this repo"

OpenClaw will route all model calls through Squish.

Qwen2.5-7B INT3 is the recommended model for OpenClaw tool-use workflows:

Model RAM TTFT Tool use
qwen2.5:7b INT4 ~5.5 GB < 600 ms ✓ Reliable
qwen2.5:7b INT4 ~2.5 GB < 300 ms ✓ Good
qwen2.5:1.5b INT4 ~1.0 GB < 200 ms ~ Limited

Anything below 4B parameters has unreliable structured JSON / tool-call output. For multi-step agent loops requiring consistent tool-use, use qwen2.5:7b or larger.

Using a Custom Port

squish run qwen2.5:7b

{
  "agent": {
    "openaiBaseUrl": "http://localhost:11435/v1"
  }
}

Using a Private API Key

squish serve qwen2.5:7b --api-key mysecretkey

{
  "agent": {
    "openaiBaseUrl": "http://localhost:11435/v1",
    "apiKey": "mysecretkey"
  }
}

Streaming

Squish supports server-sent events (SSE) streaming ("stream": true). OpenClaw enables streaming by default and will receive tokens incrementally.

Troubleshooting

"Connection refused": Squish is not running. Start it with squish serve <model>.

"model not found": The model ID in your OpenClaw config does not match the model Squish loaded. Use curl http://localhost:11435/v1/models to see the available model IDs, then update your config.

Slow first response: On the first run after compression, Squish builds an optimised weight cache. Subsequent starts load in 3–5 seconds.

Tool calls return malformed JSON: This is a model-capability issue, not a Squish or OpenClaw bug. Switch to qwen2.5:7b or a larger model. See Model Capability Reality Checks.