mma-mcp

Chinese / 中文版

A Model Context Protocol (MCP) server that wraps a local Wolfram Engine, enabling AI assistants (Claude, ChatGPT, etc.) to perform symbolic math, numerical analysis, and data visualization via Wolfram Language.

Disclaimer: This is an unofficial, independent, personal project. It is not affiliated with, sponsored by, endorsed by, or certified by Wolfram Research, Inc. "Wolfram", "Wolfram Language", "Wolfram Engine", "Mathematica", and related marks are trademarks of Wolfram Research.

This software does not include any Wolfram Engine / Mathematica binaries, activation keys, license files, or other proprietary materials. Users must independently obtain and properly license their own copy of the Wolfram Engine or Mathematica in accordance with Wolfram's licensing terms.

The sole purpose of this project is to allow a licensed individual to invoke their own, locally-installed Wolfram kernel through AI assistants on their own machine, within the scope permitted by their license. Redistribution of Wolfram Engine access to third parties is not an intended use case and may violate Wolfram's licensing terms.

Features

• MCP Tools: evaluate (text) and evaluate_image (PNG, experimental) — all Wolfram Language capabilities through two universal tools
• Transports: stdio (local) and Streamable HTTP
• Security: Pre-kernel expression filtering with blacklist/whitelist modes and 29 capability groups
• Client RBAC: Per-client credentials, per-role tool and security policy control — for isolating different AI clients on the same machine
• OAuth 2.1: Authorization server for web-based MCP clients (Claude.ai, ChatGPT)
• Config-driven: Single TOML file controls all behavior

Prerequisites

• Python 3.11+
• Wolfram Engine or Mathematica (properly licensed)
• uv package manager

Quick Start

# Clone and install
git clone https://github.com/siqiliu-tsinghua/mma-mcp.git
cd mma-mcp
uv sync

# Graphics export dependencies (headless servers only — desktops already have these)
sudo apt-get install -y libfontconfig1 libgl1 libasound2t64 libxkbcommon0 libegl1

# Generate default config
uv run mma-mcp init

# Generate security group files (requires Wolfram kernel, ~1 min)
uv run mma-mcp setup

# Start server (stdio, for local MCP clients)
uv run mma-mcp serve

Client Configuration

Claude Code / VS Code (stdio)

Add to your .mcp.json:

{
  "mcpServers": {
    "mma-mcp": {
      "command": "uv",
      "args": ["--directory", "/path/to/mma-mcp", "run", "mma-mcp"]
    }
  }
}

Claude Desktop (stdio)

Add to your claude_desktop_config.json (Settings -> Developer -> Edit Config):

{
  "mcpServers": {
    "mma-mcp": {
      "command": "/path/to/mma-mcp/.venv/bin/mma-mcp"
    }
  }
}

On macOS/Linux, find the config at ~/Library/Application Support/Claude/claude_desktop_config.json or ~/.config/Claude/claude_desktop_config.json.

HTTP Transport

uv run mma-mcp serve --transport http --host 127.0.0.1 --port 8000

Configuration

All settings live in mma_mcp.toml (or pyproject.toml under [tool.mma-mcp]).

uv run mma-mcp init  # generates mma_mcp.toml with comments

Key sections:

Section	Description
[kernel]	Wolfram kernel path, timeout, output format
[server]	Transport mode, host, port
[security]	Blacklist/whitelist mode, capability groups
[tools]	Which MCP tools to expose
[tls]	Domain and DNS provider for HTTPS (Caddy)
[auth]	Client identity and role-based access control

Security

Expressions are filtered before reaching the Wolfram kernel. Symbols are extracted via regex and checked against the active policy.

Blacklist mode (default): blocks dangerous groups (system_exec, file I/O, networking, dynamic eval).

Whitelist mode: only allows symbols from explicitly enabled groups.

29 capability groups (22 safe + 7 dangerous) cover ~6000 Wolfram Language symbols. Regenerate from your local kernel:

uv run mma-mcp setup          # required after cloning (generates from your local kernel)
uv run mma-mcp setup --force   # force regeneration (e.g., after Wolfram Engine upgrade)

Client Identity & Roles

When using HTTP transport, you can configure per-client credentials and roles to isolate different AI clients (e.g., Claude and ChatGPT) connecting to the same kernel:

# Generate password hash
uv run mma-mcp hash-password

# Generate TOML snippet for a new client
uv run mma-mcp add-client alice --role admin

Each client is bound to a role that controls which tools it can access, which Wolfram symbols it can use, and resource limits (timeout, result size). Concurrent clients are isolated via a kernel worker pool — each tool call runs in an exclusive kernel process with a temporary WL context.

See the [auth] section in mma_mcp.toml for configuration details.

Development

# Run tests
uv run pytest tests/ -v

# Inspect MCP tools interactively
uv run mcp dev src/mma_mcp/server.py

CLI Commands

Command	Description
mma-mcp serve	Start the MCP server (default)
mma-mcp init	Generate default mma_mcp.toml
mma-mcp setup	Generate security group JSONs from local kernel
mma-mcp caddyfile	Generate Caddyfile for HTTPS
mma-mcp hash-password	Hash a password for config
mma-mcp add-client	Generate TOML snippet for a new AI client

Client Compatibility

Client	Long computations	Notes
Claude.ai	✔ Supported	Sends progressToken; server heartbeat keeps connection alive
ChatGPT	✘ May timeout	Does not send progressToken; has a hard timeout (~60s) independent of server heartbeat
Claude Desktop / Claude Code	Not tested	Local stdio transport

License

MIT — applies only to the code in this repository. Use of Wolfram Engine / Mathematica is governed by Wolfram Research's own license terms.