aiming-lab/MetaClaw

MetaClaw turns live conversations into continuous training data — no GPU cluster, no offline datasets. You talk to your agent; it learns and evolves. The repo wraps your model behind an OpenAI-compatible API, intercepts interactions from OpenClaw, scores each turn with a reward model (e.g. GPT-5.2), and runs online LoRA fine-tuning via Tinker's cloud API. Updated weights are hot-swapped into production with zero downtime. If you've ever wanted an agent that gets better from real usage instead of static prompt engineering, MetaClaw is built for that. It supports two learning modes: GRPO (RL from implicit feedback) and On-Policy Distillation (OPD), where a teacher model supplies per-token log-probabilities so a smaller student matches the teacher's distribution. Skill injection pulls the most relevant skill instructions into the system prompt at every turn; with skill evolution enabled, when the agent fails, an LLM analyzes the trajectory and generates new skills automatically. Serving, reward modeling, and training are fully decoupled — the agent keeps responding in real time while optimization runs asynchronously. Stack: Python, FastAPI, Tinker SDK, Kimi-2.5 (~200B MoE) or Qwen3-4B. Quick start: install deps, point OpenClaw at the MetaClaw proxy, set TINKER_API_KEY, run the conversation RL script. From there you just chat; MetaClaw collects turns, scores them, and trains. ~250 stars, actively maintained (March 2026), from aiming-lab with citations and clear config.