← Back to projects

Erdős Navigator

active

Agent navigator that turns Claude Code into a theorem proving assistant for 1,179 of Erdős's math problems.

Python FastAPI SQLite Claude Code Lean 4 Program Synthesis

Erdős Navigator

A toolkit that makes all 1,179 of Paul Erdős’s math problems programmatically accessible to coding agents. REST API, CLI, Python SDK, and a set of Claude Code skills sitting on top of a SQLite database.

Why

Most people give their coding agent a blank environment and expect it to figure everything out from training data. That’s like dropping an RL agent into a maze with no map. The navigator gives the agent a structured environment to explore: problems it can query, difficulty signals from the community, Lean formalizations it can verify against, and a workflow that knows when to stop.

What’s in it

  • 1,179 problems with full LaTeX statements (651 still open)
  • 104 problems with monetary prizes ($37,420 in unsolved bounties)
  • Community reactions from erdosproblems.com
  • Lean 4 formalization links via Google DeepMind’s formal-conjectures repo
  • Full-text search across all problem statements
  • Agent-optimized /agent/suggest endpoint with difficulty scoring

The skills

Claude Code skills define a six-phase workflow: select, understand, pick a technique, attempt, verify, document. The verification skill grades each proof component A-to-F and asks “Would a hostile referee accept this?” The math-techniques library maps problem patterns to known approaches by domain.

I wrote about the thinking behind this in building-erdos-navigator.