What it does This repo is Numerai’s official starter kit: a handful of Jupyter notebooks that walk through dataset exploration, feature neutralization, target ensembling, and model upload. It also now installs an MCP server and advertises a one-liner where you tell an agent to “find the best neural network architecture” and let it rip.

The interesting bit The pivot is the story. Numerai used to be the poster child for artisanal quant finance—crowdsourced models, staking, burning. Now the README nudges you toward agents first and relegates the notebooks to “killing time on artisan data science.” That’s either honest evolution or a hedge against declining participation.

Key highlights

• Four Colab-ready notebooks: hello world, feature neutralization, target ensemble, barebones upload
• One-liner agent setup via curl | bash and codex exec --yolo
• MCP server integration for “open-source agent skills”
• Official tournament entry point with Discord support
• 1,145 stars, suggesting a community that predates the agent pivot

Caveats

• The agent workflow is barely documented; the README shows a single provocative command and stops
• No code visible for the MCP server itself—you’re curling a remote install script blind
• “–yolo” is in the official docs, which tells you something about the expected failure mode

Verdict Worth a look if you’re curious how quant tournaments are adapting to the agent era, or if you want pre-built Numerai notebooks. Skip if you expected rigorous documentation or a fully open agent stack—you’re trusting a bash pipe and a prayer.