What is Bagel?
Bagel, from Extelligence, is an MCP server that lets an agent read robotics, drone and IoT data and answer questions about it. It loads ROS 1 and ROS 2 bags, MCAP, PX4, ArduPilot and Betaflight flight logs, CAN captures with a DBC file, ASAM MDF4, CSV, JSON and Parquet, and reads live data from rosbridge, MQTT, Postgres and InfluxDB. Every calculation over message data is a DuckDB SQL query that Bagel shows you. It also builds data-reduction pipelines from a plain description, such as keeping 10 seconds around every hard brake, and exports to Rerun, PlotJuggler, Lichtblick and LeRobot.
It is on PyPI as bagel-mcp, Apache-2.0 licensed and written in Python 3.10 to 3.12. Version 2.5.6 reached PyPI on 15 November 2026; the repository has about 407 stars.
How to install Bagel
claude mcp add bagel -- uvx --from "bagel-mcp[ros]" bagel-mcp --transport stdioIs Bagel worth it?
Bagel's best decision is that the model never does the maths. It writes a DuckDB query over the loaded topic, Bagel runs it and returns the query alongside the answer, so a claim about, say, when the IMU ran hot can be checked by anyone who reads SQL. One prompt, summarise this bag, works the same across formats that normally need separate viewers: FlightPlot for PX4, MAVExplorer for ArduPilot, Blackbox Explorer for Betaflight.
It cannot move a robot. In our reading of the code, the live rosbridge and MQTT connections subscribe and never publish, and the README says the LLM stays in front of your logs, never in the control loop. Since 2.5.5 the model's SQL runs on a separate DuckDB connection with file, network and extension access switched off and the configuration locked, and only a single SELECT is accepted. The pipelines can still write files and upload to S3, GCS or Azure with your ambient credentials, so scope those credentials to the target buckets.
The catch: the HTTP endpoint has no authentication. The SECURITY.md file says to treat it like a database socket. The pip default binds 127.0.0.1:8000 and the Compose files publish ports on localhost only, but removing that prefix exposes an unauthenticated server to every device on the network. Recorded data needs no Docker or ROS install; live robots and edge pipelines do, and only the default ROS 2 Kilted image ships for arm64; the rest are amd64 only.
Our take: for anyone who has a folder of bags and flight logs and a script called parse_bag_final_v7.py, this replaces the script with something auditable. It is an analyst, not a pilot, and that is its strength.
Where it earns its place
A PX4 drone developer after a crash on the test field
Find out whether the battery sagged or the IMU misbehaved before the drop
Point Bagel at the .ulg file with the px4 extra and ask for a health summary; it reports voltage drops and IMU outliers with timestamps and the SQL that found them. Free, and no ROS install needed.
An autonomy team at a delivery-robot company with terabytes of bags
Keep only the seconds around hard stops before uploading to the cloud
Describe the event, preview the pipeline on a sample bag, then run it at the edge so only matching windows are written and uploaded. Cuts storage and transfer bills; scope the upload credentials first.
A vehicle test engineer with CAN captures
Check brake-pressure signals across a week of test drives
Load .blf files with the DBC through the automotive extra and query decoded signals in plain language; the CAN support is marked beta, so spot-check against your usual tool.
Before you commit
- Setup
- For recorded data, claude mcp add bagel -- uvx --from "bagel-mcp[ros]" bagel-mcp --transport stdio, adding extras such as px4, ardupilot or automotive. For live rosbridge robots, clone the repo and run docker compose run --service-ports ros2-kilted, then connect over SSE at localhost:8000.
- Integrations
- ROS 1 and ROS 2 bags, MCAP, PX4, ArduPilot and Betaflight logs, CAN and MDF4, CSV, JSON and Parquet; live rosbridge, MQTT, Postgres and InfluxDB; exports to Rerun, PlotJuggler, Lichtblick and LeRobot; S3, GCS and Azure upload.
- Data and privacy
- Runs locally on your files; query results go to your model provider through the client, and a fully local model works too. The HTTP endpoint has no authentication and defaults to 127.0.0.1:8000. Model-written SQL runs sandboxed without file or network access.
- Cost at scale
- Free under Apache-2.0. Costs are model tokens and any cloud storage your pipelines upload to.
- Limits
- Python 3.10 to 3.12 only; Intel Macs may fail to build cryptography and should use Docker. Live sources and fleet pipelines need Docker. CAN and MDF4 support is beta.
What’s good
- Shows the DuckDB SQL behind every calculation
- Reads ROS, PX4, ArduPilot, Betaflight, CAN and MDF4
- Sandboxed SQL with no file or network access
- Never publishes to a robot
What’s not
- HTTP endpoint has no authentication
- Live rosbridge and edge pipelines need Docker
- Docker images other than ROS 2 Kilted are amd64 only
- Pipelines use your ambient cloud credentials
Best for
Robotics, drone and automotive engineers who spend hours digging through recorded logs and want answers they can verify.
Bagel alternatives
Bagel: common questions
- Is Bagel worth it?
- A log analyst for robots that keeps the model's arithmetic honest by showing its SQL, and that moves nothing, which makes it the safest way to point an agent at hardware data. We rate it 4.1 out of 5.
- How much does Bagel cost?
- Free and open source (Apache-2.0). Cloud upload tasks bill to your own AWS, GCS or Azure account.
- What are the downsides of Bagel?
- HTTP endpoint has no authentication; live rosbridge and edge pipelines need Docker; docker images other than ROS 2 Kilted are amd64 only; pipelines use your ambient cloud credentials.
- Who is Bagel best for?
- Robotics, drone and automotive engineers who spend hours digging through recorded logs and want answers they can verify.
- How much does Bagel cost for a team or in production?
- Free under Apache-2.0. Costs are model tokens and any cloud storage your pipelines upload to.
- Is Bagel safe for company data?
- Runs locally on your files; query results go to your model provider through the client, and a fully local model works too. The HTTP endpoint has no authentication and defaults to 127.0.0.1:8000. Model-written SQL runs sandboxed without file or network access.
- What are the best Bagel alternatives?
- The closest alternatives we have reviewed are ROS MCP Server and DroneServer.
Answers checked Nov 16, 2026.