Repository navigation
AirClaw 3.0.0
Superseded by 3.0.1. In this build, an embedding model such as
nomic-embed-textcould be auto-selected as your chat model, failing every request. Use 3.0.1.
Run OpenClaw on a local model. Zero API cost.
Install
pip install git+https://github.com/nickzsche21/AirClawOr download the wheel below. PyPI still serves 2.0.0 — 3.0.0 will land there shortly.
airclaw detect # what's running locally?
airclaw start # gateway on :4096
airclaw patch # wire OpenClaw to it
airclaw doctor # diagnose a broken setupIf you used 2.x, airclaw patch never worked
It wrote agent.provider = "opencode" with hostname/port keys into ~/.openclaw/config.json. OpenClaw reads ~/.openclaw/openclaw.json and expects a models.providers block — so the old patcher wrote a shape OpenClaw ignores into a file it never opens, then printed a success message.
The package was also never in the repo: pyproject.toml declared an airclaw package and CLI entry point against a directory that was never committed, so pip install . from a clone built an empty package.
Both are fixed.
Also fixed
- Prompts are no longer truncated. 2.x capped input at 512 tokens, which for a coding agent discarded nearly the whole request.
- Streaming works. 2.x accepted
stream: trueand replied with a non-SSE body, hanging any client that asked for a stream. - Tool calling is forwarded. 2.x dropped
tools/tool_choice, so agents could not call tools at all. - Correct chat templates. 2.x hardcoded Llama-2
[INST]for every model, including Qwen, Phi-3 and Llama-3. - No home-directory scan. 2.x globbed
**/openclaw/config.jsonacross your entire home directory.
New
One stable OpenAI-compatible endpoint over whichever local server you run — Ollama, LM Studio, llama.cpp, vLLM, Jan, text-generation-webui — detected automatically. Configure OpenClaw once; swap backends underneath without touching agent config.
airclaw patch now backs up first, merges instead of overwriting, and refuses to write a config it could not parse. airclaw restore undoes it.
AirLLM stays as an opt-in mode for running a model larger than your GPU, documented as seconds-per-token rather than sold as fast.
34 tests cover the patcher, backend probing, and the gateway's streaming, tool passthrough and prompt handling.