Facebook Research has released RepoProver, a multi-agent scaffold for automatically formalizing mathematics textbooks in Lean. The system orchestrates LLM agents—sketchers, provers, and reviewers—that collaborate on a shared git repository to translate textbook content into formal proofs. This code successfully formalized the graduate textbook Algebraic Combinatorics by Darij Grinberg.

The tool requires Python 3.10+ and operates on a Lean project with Mathlib. Users must set up LaTeX source files organized by chapter, create a manifest.json listing target theorems, and establish a file-system-based issue tracker. The coordinator manages agent workflows through a merge queue ensuring the main branch always builds. For distributed runs, the stool launcher enables multi-node execution across SLURM clusters.

RepoProver includes analysis scripts for token usage and agent efficiency, plus a toy project for testing. The system represents a significant advance in automated theorem proving, demonstrating that large language models can effectively collaborate on complex mathematical formalization tasks. By breaking down the formalization process into specialized agent roles and managing coordination through git infrastructure, RepoProver provides a scalable framework for translating mathematical knowledge into machine-verifiable form.