Hi!
We are releasing yet another MCP server for rosbags, and an MCP Lab UI so that you can see which LLMs (closed or open) are able to properly use the available tools.
Quick demos
With Claude Desktop
With MCP Lab running locally
rosbags-mcp
The rosbags-mcp server allows LLMs to pull data from your rosbags and plot it in interactive plots with plotly. It is a pure Python package with no ROS dependencies (based on the rosbags library).
Some of the features include:
-
listing bags in a directory, and getting their info
-
searching messages (based on conditions)
-
filtering bags
We also believe that there is value in creating domain-specific tools that are useful across ROS robots. We started with basic functionality:
-
trajectory analysis (e.g., with spatial subsampling)
-
laserscan data analysis
-
logs from /rosout
-
tf tree
We also decided to add a few tools to visualize data (even though we’d love to have foxglove integration, for example!)
-
time series data plots
-
2D plots
-
laserscan polar plots
Finally, you can also return an image with an LLM (if you are using a multi-modal model).
All this is not entirely new. For example, Bagel is an amazing tool that has been posted already here (release), here (update), and here (DuckDB update). However, we started working on this before we became aware of Bagel and thought it might have a place out there! We are working on other ROS-related MCP servers and hope to release more soon. We see a lot of potential, for example, in the DuckDB integration from Bagel.
MCP Lab Web UI
Together with the rosbags MCP server, we are releasing MCP-Lab, a Web UI where you can choose OpenAI, Anthropic, or open-source models provided by Groq (for now), to test their ability to call tools in any MCP server.
You can also run MCP Lab on your robot and access it remotely.
Benchmarking
We will release a short technical paper. In the meantime, here’s a preview of selected queries and whether different LLMs are able to use the MCP server properly or not:
Any feedback is welcome!
The MCP server and MCP Lab are both projects under active development so not everything will be working perfectly.