Descriptions:
Part Time Larry builds a Claude-powered MCP (Model Context Protocol) server for real-money NFL prop bet analysis on Kalshi, putting the quantitative research from his previous video into live practice during the NFC and AFC Championship games. The system exposes three tools to Claude: a DuckDB query engine over Jonathan Becker’s 72-million-trade Kalshi dataset, an ESPN scraper for per-player season statistics, and an nfl-data-py module for conditional probability lookups (e.g., likelihood of a player scoring a second touchdown given they scored a first).
The video explains how MCP tool definitions work in Python — each tool is a function with a schema description that Claude uses to select the right data source and construct the appropriate query from a plain-English question. Asking “how often do two-touchdown props pay out?” triggers a DuckDB join across trades and markets Parquet files, returning win rates segmented by price bucket. Results showed 11–15 cent two-touchdown contracts historically resolve as wins only about 4% of the time, directly informing the decision to sell the “no” side as limit orders rather than buying emotional longshots.
All five bets placed using the system won, though the creator frames this as a small-sample intellectual experiment rather than a proven edge. The video covers the practical mechanics of registering MCP tools with the Claude client, structuring return data for natural language interpretation, and combining multiple data sources into a unified analytical assistant. Code and query examples are documented on hackingthemarkets.com.
📺 Source: Part Time Larry · Published January 31, 2026
🏷️ Format: Hands On Build
