Rosclaw

ROSClaw is a runtime that connects AI agents to robots. It sits between an agent and a physical body and enforces a governed boundary: an agent proposes an action, and the runtime decides whether it may be dispatched. The site summarizes the loop as act, verify, remember, evolve. The runtime has two lanes. The cognitive lane handles missions, tool calls, context and delegated workers, with pluggable model providers that supply reasoning but never physical authority. The physical lane handles policy and approval, scopes and limits, a daemon that mediates all dispatch, and executors that perform the bounded action and return an evidence bearing receipt. Safety controls named on the site include body binding, policy, permits, leases and an emergency stop. Verified practice, including failures, is stored as episodes and memory, and skills are evaluated before being promoted to another robot body. Agent clients such as Codex, Claude Code, Hermes and OpenClaw can connect, as can vision language action models or the project's own native agent. Backends include ROS 2, hardware MCP servers, vendor SDKs, MuJoCo, Isaac and physical robots. The site labels each component with a maturity level such as experimental, component tested or simulation verified, and notes that integration surfaces are not a claim that every agent or robot is independently verified. The project is developed at Tongji University and the Shanghai Research Institute for Intelligent Autonomous Systems. No pricing is published.