rosbridge-mcp

rosbridge-mcp is a Model Context Protocol server that connects AI agents (Claude Desktop, Cursor, VS Code, and any other MCP client) to robots running ROS 2, via the standard rosbridge v2 protocol (WebSocket + JSON). You run rosbridge_server on your robot or ROS machine; this MCP server connects to it over the network and exposes 11 tools that let the AI observe topics, inspect the ROS graph and TF tree, see through the robot's camera, publish messages, call services, and drive ROS 2 actions — no ROS installation needed on the machine running the AI client.

Architecture

+--------------------+   stdio (MCP)   +----------------+   WebSocket/JSON   +------------------+   DDS   +---------+
|  AI client         | <-------------> | rosbridge-mcp  | <----------------> | rosbridge_server | <-----> |  ROS 2  |
|  (Claude, Cursor,  |                 |  (this server) |    rosbridge v2    |  (on the robot)  |         |  graph  |
|   VS Code, ...)    |                 |                |      protocol      |                  |         |         |
+--------------------+                 +----------------+                    +------------------+         +---------+

Related MCP server: ROS2 MCP Server

Quick Start (60 seconds)

pip install git+https://github.com/hieutachi/rosbridge-mcp.git

Or, once published: pip install rosbridge-mcp (PyPI — coming soon).

Add to your MCP client config (see per-client guides below for exact file locations):

{
  "mcpServers": {
    "rosbridge": {
      "command": "rosbridge-mcp",
      "env": { "ROSBRIDGE_URL": "ws://<robot-ip>:9090" }
    }
  }
}

Then ask your agent: "What topics does the robot have?"

Choose your path

Pick the guide that matches you — each one is self-contained, you don't need to read the rest of this README first:

Tools

11 tools in total. All tools return JSON. Message and args payloads use the same JSON representation of ROS messages that rosbridge uses (field names match the .msg/.srv/.action definitions).

list_topics

List all topics with their message types. No parameters.

{"topics": [
  {"name": "/chatter", "type": "std_msgs/msg/String"},
  {"name": "/cmd_vel", "type": "geometry_msgs/msg/Twist"},
  {"name": "/scan",    "type": "sensor_msgs/msg/LaserScan"}
]}

list_nodes

List all running nodes. No parameters.

{"nodes": ["/talker", "/listener", "/rosapi"]}

list_services

List all available services. No parameters.

{"services": ["/rosapi/topics", "/rosapi/nodes", "/reset_odometry"]}

get_topic_snapshot

Subscribe to a topic, collect messages, unsubscribe. Parameters: topic (required), count (default 1), timeout seconds (default 5.0), msg_type (optional, usually auto-detected by rosbridge).

Input: {"topic": "/chatter", "count": 2, "timeout": 3.0}

{"topic": "/chatter", "requested": 2, "received": 2,
 "messages": [{"data": "Hello World: 41"}, {"data": "Hello World: 42"}],
 "timed_out": false}

If the topic is silent, received is less than requested and timed_out is true — the tool never hangs longer than timeout.

publish_message (mutating)

Advertise a topic and publish one JSON message. Parameters: topic, msg_type (full ROS 2 type, e.g. geometry_msgs/msg/Twist), message (JSON object matching the type).

Input:

{"topic": "/cmd_vel", "msg_type": "geometry_msgs/msg/Twist",
 "message": {"linear": {"x": 0.1, "y": 0.0, "z": 0.0},
             "angular": {"x": 0.0, "y": 0.0, "z": 0.2}}}

Output: {"published": true, "topic": "/cmd_vel", "type": "geometry_msgs/msg/Twist"}

call_service (mutating)

Call any ROS service. Parameters: service (required), args (JSON object, default {}), timeout seconds (default 10.0).

Input: {"service": "/rosapi/topic_type", "args": {"topic": "/scan"}}

{"service": "/rosapi/topic_type", "success": true,
 "values": {"type": "sensor_msgs/msg/LaserScan"}}

On failure the tool returns {"success": false, "error": "..."} instead of raising.

send_action_goal (mutating)

Send a goal to a ROS 2 action server (navigation, arm motion, ...). Parameters: action_name, action_type (full type with /action/, e.g. nav2_msgs/action/NavigateToPose), goal (JSON object, default {}), timeout seconds (default 30, clamped to ≤ 120), wait_for_result (default true).

Input: {"action_name": "/fibonacci", "action_type": "test_msgs/action/Fibonacci", "goal": {"order": 5}}

{"action": "/fibonacci", "goal_id": "send_action_goal:7", "success": true,
 "status": 4, "status_text": "succeeded",
 "values": {"sequence": [0, 1, 1, 2, 3, 5]},
 "last_feedback": {"partial_sequence": [0, 1, 1, 2, 3]}}

With wait_for_result: false the tool returns {"goal_id": ..., "result_pending": true} immediately — pass that goal_id to cancel_action_goal to stop the goal later. Requires a rosbridge_suite version with ROS 2 action support; against an older rosbridge the tool returns an error advising an upgrade instead of hanging.

cancel_action_goal (mutating)

Cancel a previously sent action goal. Parameters: action_name, goal_id (from send_action_goal).

Output: {"cancel_sent": true, "action": "/navigate_to_pose", "goal_id": "send_action_goal:7"}

get_tf_tree

Snapshot the robot's TF (coordinate transform) tree by listening briefly to /tf and /tf_static. Parameters: timeout seconds (default 2.0, clamped to ≤ 10).

{"frame_count": 3,
 "frames": {
   "base_link": {"parent": "odom", "translation": {"x": 1.0, "y": 0.0, "z": 0.0},
                  "rotation": {"x": 0, "y": 0, "z": 0, "w": 1}, "source": "dynamic"},
   "laser":     {"parent": "base_link", "...": "...", "source": "static"}},
 "tree": {"odom": ["base_link"], "base_link": ["laser"]},
 "roots": ["odom"]}

get_camera_image

Grab one frame from a camera topic as base64, so vision-capable models can see what the robot sees. Parameters: topic (prefer a sensor_msgs/msg/CompressedImage topic, e.g. /camera/image_raw/compressed), timeout seconds (default 5.0, clamped to ≤ 30).

Output: {"topic": ..., "format": "jpeg", "data_base64": "...", "size_bytes": 51234} (raw Image topics additionally return width/height/encoding). Frames over 4 MB are not returned — the tool responds with metadata plus an error suggesting a compressed topic.

get_connection_status

Report connection state and readonly mode. No parameters.

{"url": "ws://localhost:9090", "connected": true,
 "connected_since_unix": 1786903978.1, "active_subscriptions": [],
 "pending_service_calls": 0, "readonly": false}

Example conversation

You: What is the robot seeing right now?

Agent: (calls list_topics, finds /scan of type sensor_msgs/msg/LaserScan, then calls get_topic_snapshot with {"topic": "/scan", "count": 1}) The laser scanner reports 360 range readings. The closest obstacle is about 0.4 m away at roughly 90° to the left; the space directly ahead is clear for at least 2.5 m.

You: OK, drive forward slowly for a moment.

Agent: (calls publish_message with {"topic": "/cmd_vel", "msg_type": "geometry_msgs/msg/Twist", "message": {"linear": {"x": 0.1}, "angular": {"z": 0.0}}}) Published a 0.1 m/s forward velocity command. Tell me when to stop and I'll publish zero velocity.

For vision & embodied AI

Two of the read-only tools exist specifically to ground vision-language models in the robot's physical reality:

• get_camera_image returns a real camera frame as base64 — a vision-capable model (Claude, GPT-4o, or a VLA policy front-end) can literally look through the robot's camera before deciding what to do.

• get_tf_tree gives the model the robot's spatial skeleton — which frames exist (map, odom, base_link, camera, gripper) and how they are positioned relative to each other.

Combined with get_topic_snapshot (lidar, odometry, joint states) and send_action_goal (navigation, manipulation), this covers the observe → reason → act loop that vision-and-action agents need, over a plain WebSocket, with no ROS installation on the model side. Both perception tools work in readonly mode, so you can run a "look but don't touch" agent safely.

Configuration

Safety

Letting a language model publish /cmd_vel to a physical robot is a real risk. Set ROSBRIDGE_MCP_READONLY=true to run in read-only mode: publish_message, send_action_goal, and cancel_action_goal are rejected, and call_service only permits a fixed allowlist of known read-only /rosapi introspection services (topics, nodes, services, types, get_param, get_time, ...) — anything not on the list, including unknown future /rosapi services, is rejected. The read-only perception tools (get_topic_snapshot, get_tf_tree, get_camera_image) keep working. We strongly recommend starting in read-only mode with real hardware — see the full real-robot safety checklist and the deployment security model in SECURITY.md.

Privacy & legal

No telemetry, no data collection. Audited (2026-08): the only network connection this package ever opens is the WebSocket to the ROSBRIDGE_URL you configure — there are no analytics, no phone-home, no crash reporting, no hidden HTTP calls, and the code contains no logging of message contents to disk. The bundled mock server binds to 127.0.0.1 only. Robot data returned by tools goes exclusively to your MCP client (which forwards it to the LLM you chose — that part is under your control, not ours).

License compliance. All runtime and transitive dependencies carry licenses compatible with this project's MIT license — direct: fastmcp (Apache-2.0), websockets (BSD-3-Clause); key transitive: mcp (MIT), pydantic (MIT), starlette (BSD-3-Clause), httpx (BSD-3-Clause), anyio (MIT), cryptography (Apache-2.0/BSD-3). One transitive dependency, certifi, is MPL-2.0 — a file-level copyleft that only applies to modifications of certifi's own files and is compatible with MIT use and redistribution. No GPL/AGPL/proprietary code anywhere in the dependency tree, and all code in this repository is original work written for this project.

FAQ

Do I need ROS installed where the AI client runs? No. Only Python 3.10+. ROS and rosbridge run on the robot (or in Docker, or in a simulator); this server talks to them over WebSocket.

Does it work with ROS 1? The rosbridge v2 protocol is the same, so basic operations work against a ROS 1 rosbridge_server too — use ROS 1 type names (std_msgs/String). Only ROS 2 is tested in CI.

The agent says it cannot connect. Check that rosbridge is running (ros2 launch rosbridge_server rosbridge_websocket_launch.xml), that ROSBRIDGE_URL points at the right host/port, and that port 9090 is reachable (firewall). Each guide in docs/ has a troubleshooting section.

Can I try it without any robot or simulator? Yes — python -m rosbridge_mcp.mock_server 9090 starts a fake rosbridge with canned topics, then point ROSBRIDGE_URL at ws://localhost:9090.

Is my data sent anywhere? The server only connects to the ROSBRIDGE_URL you configure. Topic data is returned to your MCP client, which forwards it to whatever LLM you use — treat sensor data accordingly.

Roadmap

Staged plan with per-stage goals, deliverables, and the resources each stage needs: see ROADMAP.md. Highlights: v0.2 action client + TF + camera snapshots (done in v0.2.0), v0.3 HTTP transport + Docker image + rosbridge auth/TLS, v0.4 multi-robot fleets + MCP resources (URDF/map), v1.0 stable API + official MCP registry listing + Gazebo/Isaac Sim examples.

Support this project

rosbridge-mcp is built and maintained by one person, part-time, in its early stage. What exists today is real and tested: 11 tools covering topics, services, ROS 2 actions, TF, and camera snapshots; 43 automated tests running in CI on every commit; per-scenario documentation for 5 user paths; a readonly safety mode with a service allowlist; and an audited zero-telemetry codebase.

What the roadmap needs to become real, honestly stated:

• v0.3 (deployment & security): part-time development weeks, a small cloud VM or self-hosted runner for Docker image builds, and — most importantly — a security-minded reviewer for the rosbridge auth/TLS layer.

• v0.4 (fleets): access to 2+ simultaneously running robots or simulator instances, and design feedback from a real robotics lab (looking for an academic or industrial pilot partner).

• v1.0 (stability & ecosystem): sustained maintainer time (2 days/week for a quarter), one RTX-class GPU workstation for Isaac Sim validation — the main hardware ask of the whole roadmap — and optionally a low-cost robot ($1–3k) for hardware-in-the-loop CI.

How you can help, in increasing order of effort:

1. Star the repo — visibility genuinely helps an early project get contributors.

2. Try it on your robot or simulator and open an issue with your ROS distro + rosbridge version — compatibility reports are the cheapest way to make this robust.

3. Contribute a PR — docs/development.md explains the codebase in 10 minutes, and every roadmap item is claimable.

4. Sponsor or partner — if your lab or company can offer simulator time, hardware, a GPU workstation, or funded development time, reach out via github.com/hieutachi.

Related resources

If you are getting into robotics, the Robotics RL & UAV ebook is a companion learning resource by the author covering reinforcement learning and UAV robotics.

Contributing

Contributions are welcome! See CONTRIBUTING.md and the development guide. Please sign off your commits (DCO).

License

MIT — see LICENSE. Dependency licenses are permissive and compatible: fastmcp (Apache-2.0), websockets (BSD-3-Clause). No GPL/AGPL dependencies.

Tóm tắt tiếng Việt

rosbridge-mcp là một MCP server cầu nối giữa AI agent (Claude Desktop, Cursor, VS Code...) và robot chạy ROS 2 thông qua giao thức rosbridge (WebSocket + JSON). Không cần cài ROS trên máy chạy AI client.

Tài liệu được chia theo từng kịch bản — chọn đúng hướng dẫn cho bạn trong thư mục docs/:

• Dùng Claude Desktop — cấu hình JSON từng bước trên Windows/macOS/Linux

• Dùng Cursor / VS Code — cấu hình mcp.json trong editor

• Chưa có robot — chạy thử với Docker (ros:humble + rosbridge) hoặc TurtleBot3/Gazebo, hoặc mock server đi kèm

• Có robot thật — checklist an toàn: bật ROSBRIDGE_MCP_READONLY=true trước, đọc /odom, /scan để hiểu robot rồi mới mở quyền publish /cmd_vel

• Developer — kiến trúc code, cách thêm tool mới, chạy test với mock (không cần ROS)

11 tool: list_topics, list_nodes, list_services, get_topic_snapshot, publish_message, call_service, send_action_goal, cancel_action_goal, get_tf_tree, get_camera_image, get_connection_status. Bật ROSBRIDGE_MCP_READONLY=true để chặn mọi thao tác ghi (publish, action) khi làm việc với robot thật — các tool đọc (TF, camera, topic) vẫn hoạt động bình thường.

Tài liệu học kèm theo của tác giả: Robotics RL & UAV ebook — ebook về học tăng cường (reinforcement learning) và robot UAV.