Repository navigation
Run OpenClaw on a local model. Zero API cost.
pip install git+https://github.com/nickzsche21/AirClawPyPI still serves 2.0.0 — install from git or grab a wheel below.
Fixed since 3.0.0
An embedding model could be auto-selected as your chat model.
Ollama's /v1/models lists every model you have pulled, embedding and rerank models included. AirClaw took the first entry, so anyone with a common embedding model such as nomic-embed-text pulled could have it chosen as the chat model — and every request would fail with a confusing error from upstream.
Default selection now skips embedding, rerank, audio and image models, prefers instruction-tuned ones (coder, instruct, -it, chat), and returns a 503 naming the real problem when a backend has nothing that can hold a conversation:
ollama has no chat-capable model — it reports only nomic-embed-text, bge-large.
Those look like embedding or rerank models. Pull a chat model
(e.g. `ollama pull qwen2.5-coder:7b`), or name one with --model.
Everything in 3.0.0 is included — the patcher that actually works, streaming, tool passthrough, no prompt truncation, correct chat templates. See CHANGELOG.md.
41 tests, no GPU or network needed.