math-logic-mcp

MCP server that gives small LLMs verified symbolic-math & logic tools.

Small language models (Mistral, Llama, Phi, Gemma) struggle with multi-step math and formal logic. Instead of fine-tuning, give them tools. This project exposes a set of verified math and logic solvers via the Model Context Protocol (MCP), so any MCP-compatible LLM can call them as functions.

Features

Every result includes proof steps and verification — the LLM gets a machine-checked answer, not a guess.

Quick Start

Install (minimal — arithmetic only)

pip install -e .

Install (full — all solvers)

pip install -e ".[full]"

Run the MCP server

# stdio transport (for Claude Desktop, Cursor, etc.)
math-logic-mcp

# HTTP/SSE transport (for remote clients)
math-logic-mcp --http

Configure in Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "math-logic": {
      "command": "math-logic-mcp"
    }
  }
}

Configure in Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "math-logic": {
      "command": "math-logic-mcp"
    }
  }
}

Use as a Python Library

from math_logic import MathLogicEngine

engine = MathLogicEngine()

# Arithmetic (always available)
result = engine.solve("compute 2 + 3 * 4")
print(result.solutions)  # ['14']

# Algebra (requires sympy)
result = engine.solve("Solve x^2 - 4 = 0")
print(result.solutions)  # ['x = -2', 'x = 2']

# Logic (requires z3-solver)
result = engine.solve('Check satisfiability of "p and q"')
print(result.solutions)  # ['Satisfiable: p=True, q=True']

Architecture

LLM ─── MCP Protocol ──▶ mcp_server.py
                              │
                          engine.py  (router → solver → result)
                              │
              ┌───────────────┼───────────────┐
              ▼               ▼               ▼
     ArithmeticSolver   SymPySolver      Z3Solver
      (zero deps)       (pip: sympy)    (pip: z3-solver)

The router classifies each problem by regex patterns and routes to the best available solver. Solvers are loaded lazily — if SymPy isn't installed, algebra problems gracefully report the missing dependency.

Installation Extras

pip install -e ".[sympy]"     # algebra + calculus
pip install -e ".[z3]"        # logic
pip install -e ".[full]"      # everything
pip install -e ".[full,dev]"  # everything + dev tools

Development

git clone https://github.com/user/math-logic-mcp.git
cd math-logic-mcp
pip install -e ".[full,dev]"
pytest tests/ -v

Docker

docker build -t math-logic-mcp .
docker run -p 8080:8080 math-logic-mcp

License

Apache 2.0 — see LICENSE.