Winners of the Metabase AI Hackathon

We ran the Metabase AI Hackathon to celebrate every AI feature in Metabase going open source.

The submissions were genuinely fun to go through: analytics agents, new products, an entire API with machine learning, a couple of things that made us laugh, a couple that made us think “oh, that’s actually clever.” Picking two was harder than we expected.

Here are the winners:

Meta Chess, by Marat Surmashev

A live dashboard where Claude and Codex play chess against each other, with Metabase as the entire game platform, not a passive viewer.

This one uses two of our newest features in tandem. The agents sync through the MCP server: read the board state, check whose turn it is, and see the opponent’s last move. What spectators watching the dashboard see and what the agents see is the same source of truth.

The dashboard itself is built with file-based development: the live chessboard, the last-move highlight, and the move history are all Metabase YAML cards shipped through the serialization v2 API.

We picked this one because it shows off two of our newest features doing something neither was designed for. The MCP server isn’t just feeding one agent, it’s the message bus between two. File-based development isn’t just for version control, it’s how the whole dashboard ships as code. Turns out a Metabase dashboard can be a chessboard, a referee, and a livestream all at once. Who knew.

See the project · GitHub repo

Claude’s Advice, by Owais Mumtaz

A nightly Cowork task that reads the last 60 days of Owais’s personal fitness data via our MCP server, reviews his goals, and writes back to his dashboard as a chart titled “Claude’s Advice.”

Calories, macros, workouts, weight — all facts and dimensions in a personal database, with a semantic layer and predefined metrics. When Metabase 60 dropped MCP support, Owais realized Claude could use what he’d already built. The data was already mapped. Based on his current plateau, the chart told him to refeed this weekend: 3,000 calories, 400g carbs, hit legs hard, check the scale Wednesday.

We picked this one because it’s small, useful, and real. No big architecture diagram, no fancy multi-agent setup. Just one nightly task, one trusted semantic layer, and one chart that tells Owais what to do tomorrow. The kind of project that a lot of people could build for themselves this weekend… and probably should.

Watch Owais’s walking through the project

Two mechanical keyboards are on their way to Marat and Owais. Thanks to everyone who participated.

If you haven’t tried AI in Metabase yet, you don’t need a hackathon to start, go build something and tag us, we read everything