MetaClaw

MetaClaw is an open-source continual-learning layer for OpenClaw agents. It sits between OpenClaw and the underlying model as an OpenAI-compatible proxy, captures live user-agent conversations, scores outcomes, and uses those interactions to improve the serving policy over time. The project combines online fine-tuning, skill retrieval, and optional skill evolution so behavior can improve both immediately through prompt-time skill injection and more durably through background training. Its architecture is explicitly decoupled: serving continues in real time while reward modeling and optimization run in parallel, and updated weights are hot-swapped into production without restarting the service. The repository also emphasizes lower operational friction than traditional RL systems by offloading training to Tinker cloud instead of requiring a dedicated local GPU cluster. For OpenClaw operators experimenting with adaptive agents, MetaClaw is best understood as infrastructure for post-deployment learning rather than a standalone end-user product or dashboard.

Use this tool if you:\n- Want an OpenClaw agent to learn continuously from real user conversations instead of static offline datasets.\n- Need an OpenAI-compatible proxy layer that can intercept traffic, score interactions, and improve policies over time.\n- Want immediate behavior gains from retrieved skill injection while longer-term training happens in the background.\n- Need continual improvement without pausing or restarting the live serving path.\n- Are comfortable operating an experimental training stack and evaluating agent behavior over repeated interaction cycles.