It appears the browse tool was unable to access the provided URLs. I will proceed with generating the brief based on the information directly available in the initial prompt, as I cannot fetch external details.

Numina-Lean-Agent is an AI agent designed for formal theorem proving tasks, built upon Claude Code. It targets researchers and practitioners in formal mathematics, offering a system capable of tackling complex problems. The agent has demonstrated its efficacy by proving all 12 problems from Putnam 2025 and completing a formalization of Effective Brascamp-Lieb inequalities, providing a powerful tool for advancing formal mathematical reasoning.

How It Works

The system leverages Claude Code and a suite of modular CLI skills, including code-transform, llm, search, sorrifier, and verification. It operates within the Lean 4 formal proof management system, requiring Lean code to reside within a buildable Lean project structure managed by lake. The agent employs an "autosearch coordinator prompt" and executes in multiple rounds to refine proofs and explore solutions.

Quick Start & Requirements

Installation involves cloning the repository, running a setup script (./setup.sh YOUR_PROJECT_NAME) within the tutorial directory, and then setting up the Python environment using uv python install and uv sync. Activation requires source .venv/bin/activate. The agent requires API keys for Claude, Gemini, OpenAI, Axle, and Leandex, alongside specific Lean 4 project setup. A quick example run is provided via example_run.sh. Detailed setup and usage guides are available.

Highlighted Details

• Successfully proved all 12 problems from Putnam 2025.
• Completed a paper-level formalization of Effective Brascamp-Lieb inequalities.
• Modular architecture with distinct CLI skills for various reasoning tasks.
• Supports multiple LLM backends (Claude, Gemini, OpenAI) for flexibility.

Maintenance & Community

The project is associated with authors listed in the Numina-Lean-Agent paper. Related projects include lean4-skills and Leandex. Specific community channels (e.g., Discord, Slack) or a public roadmap are not detailed in the provided README.

Licensing & Compatibility

The project is released under the MIT License, which generally permits broad use, modification, and distribution, including for commercial purposes, with standard attribution requirements.

Limitations & Caveats

A critical requirement is that all .lean files or directories targeted by the agent must reside within a properly configured and buildable Lean project; standalone files outside such projects will cause errors. The setup necessitates configuring multiple external API keys, which can be a barrier to entry. The automatic setup script may require manual intervention, with a detailed guide provided for troubleshooting.