ROS MCP Server

ROS MCP Server puts a language model on the other end of a robot's message bus. It connects a Model Context Protocol client to a ROS 1 or ROS 2 system through the standard rosbridge node, which means the robot itself needs no modification — no new nodes to write, no wrapper API to maintain, nothing recompiled. If rosbridge is already running, the robot is already addressable.

Communication is bidirectional, and that is what makes it more than a remote control. The assistant can publish to topics, call services, send action goals and set parameters, but it can equally subscribe to sensor streams and read back state — so it can act, observe the result, and adjust. That closes the loop for the debugging work robotics engineers actually spend their days on: why did the arm stop mid-trajectory, which node dropped the transform, what is odometry reporting right now.

The server also handles discovery rather than assuming prior knowledge of your robot. It guides the model to enumerate available topics, services and actions and to inspect their message types before composing a call, which is the difference between a usable integration and one that guesses at message field names. Both ROS 1 distros and ROS 2 (Humble, Jazzy and others) are supported.

Install with uvx ros-mcp --transport=stdio — for Claude Code, claude mcp add ros-mcp -- uvx ros-mcp --transport=stdio. It works with any MCP-compatible client including Claude Code, Claude Desktop, Cursor, Codex CLI and Gemini CLI. Requires Python 3.10+, and the client machine must reach the robot's rosbridge over the local network or a VPN. Apache 2.0 licensed.

A safety note worth stating plainly: this grants an AI assistant authority to command physical hardware. Test against a simulator before pointing it at a real platform.