Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

ROS MCP Server

The most-used bridge from an agent to a ROS robot: full topic, service and action access through rosbridge, with no safety layer of its own.

Reviewed by ReviewRamble editors · Updated

Setup docs Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.9/5

The obvious starting point for putting an agent on a ROS robot, provided you accept that the only thing between the model and the motors is your own ROS stack.

What is ROS MCP Server?

ROS MCP Server, from the robotmcp group, connects an MCP client to a robot running ROS 1 or ROS 2 through rosbridge, the WebSocket bridge that ships with ROS. The agent can list and inspect topics, services, actions, nodes and parameters, subscribe to sensor data, read camera images, publish messages, call services and send or cancel action goals. Nothing changes in the robot's own code; you add the rosbridge node on the robot and run the server on your own machine.

It is published on PyPI as ros-mcp, Apache-2.0 licensed and written in Python. The repository has about 1,527 stars and 219 forks, released v3.1.2 on 23 September 2026 and had commits on 15 November 2026.

How to install ROS MCP Server

claude mcp add ros-mcp -- uvx ros-mcp --transport=stdio

Is ROS MCP Server worth it?

ROS MCP Server is the reference most other ROS servers are measured against, and its reach is the reason. In our reading of the code it registers 31 tools, and the agent discovers custom message, service and action types on its own, so it can drive a robot it has never seen with the right syntax. Because it talks to rosbridge rather than linking against ROS, the same server works with ROS 1 and ROS 2 and can sit on a laptop while the robot runs elsewhere.

What it can move is whatever your ROS graph lets a publisher move. publish_for_durations will send a velocity command to /cmd_vel at a set rate for a set number of seconds; call_service and send_action_goal reach grippers, arms and navigation stacks. There is no allowlist of topics, no speed or workspace limit and no confirmation step in the server. The only guard is MCP's destructiveHint annotation on tools that change state, which some clients turn into an approval prompt and others ignore. The connect_to_robot tool even lets the model change which IP address it is talking to.

The catch: the network. rosbridge opens port 9090 on the robot, and the setup guide configures no authentication for it, so anything on that network that can reach the port can publish to the robot too. The server itself defaults to stdio, but its HTTP mode has no authentication, and the ChatGPT guide puts it behind a public ngrok tunnel with the connector set to No authentication. Open issues include send_action_goal reporting success when rosbridge returned a failure, and concurrent action calls returning each other's results.

Our take: run it first against the Turtlesim Docker example or a Gazebo or Isaac Sim robot, then on hardware only with a speed limiter, a geofence or a supervisor node already in your ROS stack and a person at the e-stop. It is a capable bridge, not a safety system.

Where it earns its place

  • A field engineer at an industrial automation integrator

    Diagnose a misbehaving gripper from the technician manual

    Connect over the plant LAN and let the agent read topics, call the gripper's test services and compare readings with the manual. Free to run; restrict the session to read-only tools or approve each destructive call in the client.

  • A university robotics lab with a TurtleBot3 and a Gazebo setup

    Let students give a mobile robot plain-language navigation goals

    Run the server against the simulated robot first, then the real one with Nav2 handling path planning and speed limits. The server sends the goals; Nav2 decides how fast the robot moves.

  • A start-up prototyping a quadruped demo in NVIDIA Isaac Sim

    Script walk-throughs of a Unitree Go2 by prompt before touching hardware

    Use the bundled unitree_go2 robot specification and publish velocity commands in simulation, then reuse the same prompts on the real robot once limits are set on the robot side.

Before you commit

Setup
On the robot, install and launch rosbridge with ros2 launch rosbridge_server rosbridge_websocket_launch.xml (the launch file also starts rosapi, which the introspection tools need). On your machine, claude mcp add ros-mcp -- uvx ros-mcp --transport=stdio, then ask the agent to connect to the robot's IP.
Integrations
ROS 1 and ROS 2 (Jazzy, Humble and others) through rosbridge on port 9090; topics, services, actions, nodes, parameters and camera images. Setup guides for Claude Code, Codex CLI, Gemini CLI, Claude Desktop, ChatGPT and Cursor.
Data and privacy
Runs on your machine; topic data and camera frames go to your model provider through the client. The MCP server defaults to stdio; --transport http binds 127.0.0.1:9000 unless you pass --host 0.0.0.0, and has no authentication either way.
Cost at scale
Free under Apache-2.0. The costs are the model's tokens, which grow with image and high-rate topic reads, and the robot or simulator.
Limits
No safety envelope in the server. send_action_goal works only on ROS 2; some parameter and action tools are missing on ROS 1 and before Humble. About 24 open issues, including action results being misreported.

What’s good

  • 31 tools covering topics, services, actions and parameters
  • Works with ROS 1 and ROS 2 through rosbridge
  • Discovers custom message types without configuration
  • One-line uvx install for Claude Code and other clients

What’s not

  • No server-side limits, allowlist or confirmation
  • rosbridge port 9090 set up without authentication
  • ChatGPT guide uses a public tunnel with no auth
  • Open bugs in action result reporting

Best for

ROS developers and researchers who want an agent to inspect and drive a robot whose own ROS stack already enforces speed and safety limits.

ROS MCP Server alternatives

ROS MCP Server: common questions

Is ROS MCP Server worth it?
The obvious starting point for putting an agent on a ROS robot, provided you accept that the only thing between the model and the motors is your own ROS stack. We rate it 3.9 out of 5.
How much does ROS MCP Server cost?
Free and open source (Apache-2.0). You supply the robot or simulator and the model.
What are the downsides of ROS MCP Server?
No server-side limits, allowlist or confirmation; rosbridge port 9090 set up without authentication; chatGPT guide uses a public tunnel with no auth; open bugs in action result reporting.
Who is ROS MCP Server best for?
ROS developers and researchers who want an agent to inspect and drive a robot whose own ROS stack already enforces speed and safety limits.
How much does ROS MCP Server cost for a team or in production?
Free under Apache-2.0. The costs are the model's tokens, which grow with image and high-rate topic reads, and the robot or simulator.
Is ROS MCP Server safe for company data?
Runs on your machine; topic data and camera frames go to your model provider through the client. The MCP server defaults to stdio; --transport http binds 127.0.0.1:9000 unless you pass --host 0.0.0.0, and has no authentication either way.
What are the best ROS MCP Server alternatives?
The closest alternatives we have reviewed are Bagel, DroneServer and WLKATA MT4 control stack.

Answers checked Nov 16, 2026.