SageMath MCP Server

A Model Context Protocol server that exposes SageMath — a 100+ package computer algebra system — to LLM clients such as Claude. Where most "math MCPs" wrap a single library (SymPy, mathjs), this one subprocess-wraps the entire SageMath stack so the model has access to genuine symbolic mathematics, not arithmetic.

Why

LLMs are bad at exact arithmetic, brittle with multi-step algebra, and unreliable when symbolic manipulation matters. The right pattern is to keep the conceptual reasoning in the model and delegate mechanical computation to a tool — the same trick a person with dyscalculia uses with a calculator. SageMath gives you the broadest tool surface for that delegation: factoring, integration, eigenvalues, number theory, LaTeX rendering, and plotting all behind one interface.

Tools

All tools follow the same schema pattern (see src/schemas/common.ts).

Installation

The server needs either a local SageMath install or Docker available on the host so it can spawn a SageMath process.

Option 1: npm + local SageMath

npm install -g @justice8096/sagemath-mcp-server
# requires `sage` on PATH — see https://doc.sagemath.org/html/en/installation/
sagemath-mcp

Option 2: Docker (no local SageMath required)

docker compose up
# uses the official sagemath/sagemath image; the server runs inside the container

Option 3: from source

git clone https://github.com/justice8096/sagemath-mcp-server.git
cd sagemath-mcp-server
npm install
npm run build
npm start

Configuring an MCP client

Claude Desktop / Claude Code (stdio transport)

Add to ~/.claude.json (or the equivalent Claude Desktop config):

{
  "mcpServers": {
    "sagemath": {
      "command": "sagemath-mcp"
    }
  }
}

Remote / HTTP transport

TRANSPORT=http API_KEY=your-secret-key npm start

Then point an MCP HTTP client at http://<host>:3000/mcp with the matching API_KEY header.

Architecture

• TypeScript wrapper, runtime is Node.js ≥ 18.
• Subprocess wrapping — the wrapper code is CC0; SageMath remains GPL-3.0 and runs as a separate process, which is the canonical safe interaction model per the GNU GPL FAQ on aggregation. Embedding via FFI would propagate copyleft; subprocess execution does not.
• Dual transport — stdio (Claude Desktop, Claude Code, local MCP clients) and Streamable HTTP (remote deployment, optional API_KEY env var).
• i18n — English, Spanish, French via src/i18n.ts and locales/.
• Multi-format build — build.ts emits Claude plugin, OpenAI function-calling, n8n node, MCP server descriptor, CLI tool, and YAML prompt-library outputs from the same source manifest.

Configur

... View full README on GitHub