A friend of mine plays a lot of Arena and kept asking ChatGPT for advice, and as many of you know, it worked badly! It would confidently recommend things that didn't exist or had rotated out of Standard months ago. I thought it was a cool problem to tackle, so I built ten MCP reference modules that give Claude and ChatGPT access to real MTG data: 17Lands draft stats across all 31 color archetypes, Frank Karsten's hypergeometric mana base math, Scryfall's full card database, and the MTG Comprehensive Rules with semantic search.
The rabbit hole I fell deepest into was the draft advisor. It's an 8-axis WASPAS scoring engine — baseline win rate (Bayesian-shrunk so sparse archetypes don't produce noise), N-wise card synergy, curve fit, castability via Karsten's tables, signal openness, role composition, color commitment, and opportunity cost. Empirical winning-deck data only works for card-intrinsic axes; state-dependent axes like signal and opportunity cost need theoretical sigmoid parameters or survivorship bias destroys the differentiation.
Try it! The MCP works out of the box with MTG rules, cards, stats, and mana base. If you connect the lightweight Savecraft daemon to your Arena install, it'll watch your `Player.log` and context will automatically flow to the LLM on your specific decks and matches, enabling the draft advisor and play advisor modules (and other cool stuff).
A friend of mine plays a lot of Arena and kept asking ChatGPT for advice, and as many of you know, it worked badly! It would confidently recommend things that didn't exist or had rotated out of Standard months ago. I thought it was a cool problem to tackle, so I built ten MCP reference modules that give Claude and ChatGPT access to real MTG data: 17Lands draft stats across all 31 color archetypes, Frank Karsten's hypergeometric mana base math, Scryfall's full card database, and the MTG Comprehensive Rules with semantic search.
The rabbit hole I fell deepest into was the draft advisor. It's an 8-axis WASPAS scoring engine — baseline win rate (Bayesian-shrunk so sparse archetypes don't produce noise), N-wise card synergy, curve fit, castability via Karsten's tables, signal openness, role composition, color commitment, and opportunity cost. Empirical winning-deck data only works for card-intrinsic axes; state-dependent axes like signal and opportunity cost need theoretical sigmoid parameters or survivorship bias destroys the differentiation.
Try it! The MCP works out of the box with MTG rules, cards, stats, and mana base. If you connect the lightweight Savecraft daemon to your Arena install, it'll watch your `Player.log` and context will automatically flow to the LLM on your specific decks and matches, enabling the draft advisor and play advisor modules (and other cool stuff).
Everything is open source and Apache 2.0 @ https://github.com/joshsymonds/savecraft.gg