by nickzsche21
Run OpenClaw on a local model. Zero API cost. One OpenAI-compatible endpoint over Ollama, llama.cpp, vLLM or LM Studio — plus AirLLM for models bigger than your GPU.
# Add to your Claude Code skills
git clone https://github.com/nickzsche21/AirClawLast scanned: 8/31/2026
{
"issues": [],
"status": "PASSED",
"scannedAt": "2026-08-31T10:35:34.344Z",
"npmAuditRan": true,
"pipAuditRan": true,
"promptInjectionRan": true
}AirClaw is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by nickzsche21. Run OpenClaw on a local model. Zero API cost. One OpenAI-compatible endpoint over Ollama, llama.cpp, vLLM or LM Studio — plus AirLLM for models bigger than your GPU. It has 116 GitHub stars.
Yes. AirClaw passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.
Clone the repository with "git clone https://github.com/nickzsche21/AirClaw" and add it to your Claude Code skills directory (see the Installation section above).
AirClaw is primarily written in Python. It is open-source under nickzsche21 on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh AirClaw against similar tools.
No comments yet. Be the first to share your thoughts!
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM is an informational directory and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
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 git+https://github.com/nickzsche21/AirClaw
airclaw detect # what's running locally?
airclaw start # gateway on :4096 (leave running)
airclaw patch # wire OpenClaw to it
Install from git for now. PyPI still serves 2.0.0, whose
airclaw patchwrote a config OpenClaw ignores and then reported success. Use the command above, or grab a wheel from releases.pip install airclawbecomes the right command once 3.0.0 is published.
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.
AirClaw finds these automatically, in this order:
| Backend | Default port | Notes |
|---|---|---|
| Ollama | 11434 | Easiest. ollama pull qwen2.5-coder:7b |
| LM Studio | 1234 | GUI, lms server start |
| llama.cpp | 8080 | llama-server -m model.gguf |
| vLLM | 8000 | Fastest if you have the VRAM |
| Jan | 1337 | |
| text-generation-webui | 5000 |
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 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.
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.
| Command | What it does |
|---|---|
airclaw detect |
List running local inference servers |
airclaw start |
Start the gateway on :4096 |
airclaw patch |
Write AirClaw into the OpenClaw config |
airclaw doctor |
Diagnose the whole chain, top to bottom |
airclaw status |
Is the gateway up? |
airclaw restore |
Undo the config change |
airclaw models |
List AirLLM aliases |
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).
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:
stream: true and replied with a non-SSE
JSON body, which hangs clients that asked for a stream.tools/tool_choice entirely, so
agents could not call tools.[INST] formatting for every
model, including Qwen, Phi-3 and Llama-3, which use different templates.glob("**/openclaw/config.json") across
your entire home directory.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 --force-reinstall git+https://github.com/nickzsche21/AirClaw
airclaw restore # only if 2.x touched a config you want reverted
airclaw patch
airclaw doctor
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.
MIT