Post

Another step in getting tabular foundation models into the working knowledge of agents: @LennartPurucker built a TabArena MCP server, so an agent can query the leaderboards directly.

I did a very loose experiment by asking Claude Code about how to build an agent that makes predictions over structured data. It pulled the actual numbers and picked best accuracy/latency compromise (TabPFN-3), then wired it in as a tool the LLM calls rather than a prediction the LLM makes itself.