OpenClaw agents can now run on open models via Hugging Face or llama.cpp

OpenClaw just got a lifeline for anyone who’s been locked out of their agents. The team has published a guide for migrating OpenClaw, Pi, or Open Code agents away from proprietary APIs and onto open models — with two clear paths depending on what you care about most.

The fastest route runs through Hugging Face Inference Providers, which acts as a routing layer to open-source models hosted on external infrastructure. You grab a token from Hugging Face, run openclaw onboard --auth-choice huggingface-api-key, and paste it in. The team recommends GLM-5 for its strong Terminal Bench performance, though you can swap in any model by dropping its repo_id into your OpenClaw config file. HF PRO subscribers get $2 in monthly credits that apply to inference usage, which softens the cost for anyone already in that ecosystem.

The second option is fully local via llama.cpp — the open-source inference library built for low-resource hardware. Install it through Homebrew on Mac or winget on Windows, then fire up a server with a single command pointing at a GGUF model. Their example uses Qwen3.5-35B-A3B, which hums along nicely on 32GB of RAM. The onboarding command is a bit more involved since you’re specifying a custom base URL, model ID, and compatibility mode, but the guide lays out the exact flags you need.

What’s notable here isn’t just the technical walkthrough — it’s the framing. The guide ends with an explicit statement that you don’t need a closed hosted model to get OpenClaw working again. That’s a clear signal about where the project sees its future, and it puts pressure on the proprietary API vendors that have been the default for agent frameworks. Whether the open models can match the reliability and reasoning quality that agent workflows demand is the real question, and one the guide doesn’t pretend to answer — it just makes sure you can find out for yourself.