AirClaw

Run OpenClaw on a local model. Zero API cost.

AirClaw puts one stable OpenAI-compatible endpoint in front of whatever local
inference server you already run, then writes the OpenClaw config that points at
it. Configure OpenClaw once; swap Ollama for llama.cpp for vLLM underneath
without touching agent config again.

pip install airclaw

airclaw detect     # what's running locally?
airclaw start      # gateway on :4096  (leave running)
airclaw patch      # wire OpenClaw to it

Restart OpenClaw. That's it — your agent now runs on your own hardware.

Something not working? airclaw doctor checks all three links in the chain and
tells you which one is broken.

Backends

AirClaw finds these automatically, in this order:

Force one, or point at something else entirely:

airclaw start --backend ollama
airclaw start --backend http://192.168.1.50:8000/v1
airclaw start --model qwen2.5-coder:14b

AirLLM mode

AirLLM streams model layers off disk one at a time, which is how it fits a 70B
model into about 4GB of VRAM.

It is slow. Expect seconds per token, not tokens per second. It is a genuine
way to run a model your GPU cannot hold, and it is not a way to run an
interactive coding agent. It is opt-in for exactly that reason:

pip install 'airclaw[airllm]'
airclaw start --airllm --model coder

Aliases: 7b 8b 13b 70b qwen coder deepseek phi, or any Hugging
Face model id. airclaw models lists them. Tool calling is not available in
this mode — the gateway returns a clear 400 rather than pretending.

How it works

OpenClaw ──> AirClaw gateway :4096 ──> Ollama / llama.cpp / vLLM / LM Studio
             (stable alias                (whatever is actually running)
              "airclaw/airclaw")

airclaw patch writes a models.providers.airclaw block into your OpenClaw
config and sets agents.defaults.model.primary to airclaw/airclaw. It backs
the file up first, merges rather than overwrites, and refuses to write a config
it could not parse. airclaw restore puts the original back.

The gateway forwards streaming, tool/function calling, and sampling parameters
untouched. It does not truncate prompts.

Commands

Useful flags: --port, --host, --config, --no-default (register the
provider without making it the default model), --create (make the config file
if OpenClaw hasn't yet).

Upgrading from 2.x

If you installed AirClaw 2.x, airclaw patch did not work. 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 — and printed a success message.

Also fixed in 3.0:

• Prompts are no longer truncated. 2.x capped input at 512 tokens, which for
  a coding agent meant discarding nearly the whole request.
• Streaming works. 2.x accepted stream: true and replied with a non-SSE
  JSON body, which hangs clients that asked for a stream.
• Tool calling is forwarded. 2.x dropped tools/tool_choice entirely, so
  agents could not call tools.
• Correct chat templates. 2.x hardcoded Llama-2 [INST] formatting for every
  model, including Qwen, Phi-3 and Llama-3, which use different templates.
• No home-directory scan. 2.x ran glob("**/openclaw/config.json") across
  your entire home directory.
• The package is actually in this repo. 2.x's pyproject.toml declared a
  package and a CLI entry point against a directory that was never committed, so
  pip install . from a clone produced an empty package.

To upgrade:

pip install --upgrade airclaw
airclaw restore   # only if 2.x touched a config you want reverted
airclaw patch
airclaw doctor

Development

uv venv --python 3.12
uv pip install -e '.[dev]'
pytest

Tests run against a stub OpenAI-compatible server over a real socket — no GPU,
no model download, no network.

License

MIT