@math-mentor/agent-skills

An MCP (Model Context Protocol) extension for Claude Code that helps you deeply explore hard math papers. Drop a PDF into your project, and use slash commands to extract theorems, generate depth analyses, run interactive Q&A sessions, and more -- all calibrated to your math level and proof preferences. Official npm: https://www.npmjs.com/package/@math-mentor/agent-skills?activeTab=code

Quick Start

Prerequisites: Node.js >= 18, Claude Code

1. Install

npx @math-mentor/agent-skills init

You'll be guided through an interactive setup:

? What is your current math level?
  ❯ High School
    Undergraduate
    Graduate
...

2. Add papers

Drop any PDF into the generated papers/ directory.

3. Open Claude Code

claude

4. Verify the MCP connection

Type /mcp inside Claude Code. You should see mathmentor listed as a connected server with 5 tools.

5. Start exploring

Talk to Claude inside the project directory. The MathMentor tools are now available. You can ask claude about a paper in natural language and it will automatically recognize the tools. You can also try any of the commands belows.

The .md files with LaTex are saved under mathmentor subdirectory.

Commands

Command	Description
/extract <paper>	Extract all theorems, definitions, and algorithms from a paper
/summary <paper>	Generate a concise summary
/depth <paper>	In-depth research-question-focused analysis
/debate <paper>	Interactive Q&A tutoring session
/prerequisites <paper>	List prerequisite knowledge needed to understand the paper
/compare <paper1> <paper2>	Compare two papers side by side
/reading-graph	View your reading progress and connections between papers

mcp-demo-1.mp4

All outputs are calibrated to your configured math level and saved to mathmentor/ for future reference.

Configuration

Math Levels

Level	What it assumes	What it explains
High School	Algebra, basic calculus	Everything from scratch. Analogies, concrete examples, no proofs.
Undergraduate	Analysis, linear algebra, topology, probability	Specialized concepts. Proof sketches, LaTeX notation.
Graduate	Measure theory, functional analysis, grad-level probability	Only paper-specific constructions. Full technical detail.

Proof Preference

Independent of math level, you can configure how much proof detail you want:

Preference	What you get
intuition-only	No proofs at all. Focus on what results say and why they matter. Best for quick surveys or when you just need the big picture.
sketches	Proof sketches (3-5 sentences per major result). Captures the key trick or technique without full detail. The default.
full-proofs	Detailed proof walkthroughs with intermediate steps, key inequalities, and technique names. For deep study.

This is set during init and stored in .mathmentor/config.json. You can change it anytime by editing the file directly or re-running npx @math-mentor/agent-skills init.

Debate Mode

After extracting and analyzing a paper, enter debate mode to discuss it live. Claude reads all your existing analysis outputs -- extractions, depth analyses, and prior Q&A -- then answers your questions with full mathematical detail, appending each exchange to a persistent debate file.

mcp-demo-2.mp4

How It Works

The package provides:

• MCP Server with 5 tools (find_paper, get_config, update_reading_graph, get_reading_graph, list_artifacts)
• 7 slash commands for Claude Code
• 3 specialized agents (extraction, depth analysis, debate)
• 3 skills (paper finding, question formulation, math level calibration)
• Reading graph that tracks your progress across papers and identifies connections

License

MIT