sagebrush-mcp: computational mathematics for AI agents

An MCP server that gives an AI agent a real computer algebra system for research mathematics, with Sage syntax and exact, verified results. It is Sagebrush: Rust engines, MIT OR Apache-2.0. You don't need Sage, PARI or FLINT, or a C compiler; the wheels cover Linux, macOS and Windows.

pip install sagebrush-mcp        # or: uvx sagebrush-mcp

Claude Code: claude mcp add sagebrush -- sagebrush-mcp

Claude Desktop, Cursor and other clients: in the MCP configuration,

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

(uvx sagebrush-mcp as the command works without installing anything first.)

Tools

• sage: run Sage code in a persistent session. 2^10, 1/3 exact, R.<x> = QQ[], K.<a> = NumberField(x^3 - 11), [1..10]; variables persist between calls. It returns the printed output and the last value, and plots come back as SVG.
• python: the same session, as plain Python.
• factor, number_field, newforms: shortcuts for common questions.
• guide: what is implemented. reset: clear the session.

What the session covers:

• Integers: factoring (ECM), primality.
• Polynomials: factoring and roots over ZZ and QQ.
• Number fields: maximal orders, prime decomposition, class groups, units and regulators (assuming GRH, as Sage and PARI do).
• Exact linear algebra over ZZ and QQ: det, rref, inverse, charpoly, kernels, Hermite and Smith forms, LLL.
• Modular forms: modular symbols, newforms with LMFDB labels, Hecke operators.
• Elliptic curves: conductors, ranks, a_p.
• Plots.

The results agree with Sage. The test suite compares 3,500 lines of output with Sage itself.

Each computation has a time limit (timeout, default 60 s). When it expires, the computation is interrupted and the session keeps its variables.

This package is a thin entry point: the server is sagebrush.mcp_server in the sagebrush package, so pip install sagebrush also provides the sagebrush-mcp command.