An LLM that reads the ROS graph and drives the robot

What it does ROS-MCP-Server is a Python MCP server that connects LLM clients — Claude Desktop, Claude Code, Cursor, Codex CLI, Gemini CLI, ChatGPT — to robots running ROS 1 or ROS 2. It rides on the rosbridge node, so the robot side stays untouched: the model talks MCP on one end, ROS topics, services, actions, and parameters on the other. Communication is genuinely two-way — the LLM can drive the robot and watch what it’s doing, from sensor data to real-time state monitoring.

The interesting bit The plumbing itself is honest glue — MCP to rosbridge, nothing scandalous. The actual contribution is that the server guides the LLM to introspect the live ROS graph: enumerating topics, services, and actions, including custom types, and querying their syntax so the model can use them correctly without anyone hand-writing tool descriptions. The robot effectively teaches the model its own API at runtime.

Key highlights

• Zero robot-side code changes — just add the rosbridge node to an existing setup
• Full ROS surface: publish/subscribe, services, actions, parameters, sensor reads, state monitoring
• Runtime discovery of custom topic and service types, syntax included
• Cross-version support: ROS 2 (Jazzy, Humble, others) and ROS 1
• Demonstrated on a MOCA manipulator in Isaac Sim, a Unitree Go2 (camera interpretation included), and live debugging of an industrial robot

Caveats

• The feature list promises actions, yet “Action support” and “permissions” appear under future contributions — the README is unclear on how complete action support really is
• No permission layer yet, so an LLM gets broad robot access; demo accordingly
• The discovery mechanism’s internals aren’t documented in the README — you’re trusting the videos

Verdict If you run a ROS stack and want an LLM that can poke, probe, and pilot it — for debugging especially — this is the most direct route on offer. If you need safety interlocks or non-ROS hardware, keep walking.