OpenClaw-RL

Why we included this project

OpenClaw-RL is for people who run their own assistant or coding agent and want it to learn from real use instead of staying frozen at whatever it knew on day one. It sits in front of your deployed model as an OpenAI-compatible API, watches live multi-turn conversations, and turns the signals that follow each action, like a user re-query, a tool result, or a failing test, into training gradients. Rollout collection, reward judging, and policy updates each run on their own asynchronous loop, so the agent keeps answering requests while optimization happens in the background, on your hardware and without manual labeling. You can train with scalar process rewards, hindsight-guided on-policy distillation, or a hybrid of both, and the same setup works for terminal, GUI, software engineering, and tool-call agents. If you want to tune your own agent policy without handing conversation data to a third party, this is a practical way to do it.