SageMath MCP Server

A Model Context Protocol server that gives an LLM a sandboxed mathematical subset of SageMath — symbolic calculus, number theory, linear algebra, ODEs, plotting, combinatorics, graphs, groups, elliptic curves, and more. Each MCP session gets a dedicated Sage worker process, so variables, functions, and assumptions persist across tool calls. It ships 40 MCP tools, one of which — verify_claim — re-checks a stated result through a proof ladder and answers proved / refuted / supported / undecided with its evidence.

Caller code is deny-by-default: the full breadth of Sage mathematics is reachable, but imports, the external CAS interfaces, and the file / display / persistence primitives are not. The policy accepts 98.9% of SageMath's own 438,124 documented doctest examples (4,259 in-scope refusals, every one attributed to a named rule; 98.9% of 433,281 on the passagemath runtime) while refusing the rest — measured on every CI run (see Security).

Full manual: USAGE.md — the tools and their parameters, how code is interpreted, and the security model in depth.

Install & run

Recommended — the container image (SageMath is baked in):

docker run --rm \
  --read-only --tmpfs /tmp:rw,size=512m --tmpfs /home/sage/.sage:rw,size=256m \
  --cap-drop ALL --security-opt no-new-privileges --pids-limit 256 --memory 4g \
  -p 127.0.0.1:8314:8314 \
  ghcr.io/xbp-europe/sagemath-mcp:latest

Those flags are the hardening the server expects; the port is published on loopback deliberately — the server executes code and authenticates nobody. docker compose up --build applies the same hardening from one reviewed file. Released images are signed with Cosign and carry SLSA provenance and an SPDX SBOM as registry attestations; the PyPI files carry PEP 740 attestations. DISTRIBUTION.md shows how to verify each.

From PyPI (bring your own Sage runtime):

pip install sagemath-mcp
sagemath-mcp                                             # stdio (default)
sagemath-mcp --transport streamable-http --port 8314    # HTTP on 127.0.0.1

This needs a working SageMath on the host: sage on your PATH, or in the same environment as the server. A Sage inside a container is out of reach of a server on the host; without a Sage of your own, use the [passagemath] extra or conda-forge below, or the container image above.

From conda-forge (Linux and macOS, next to conda-forge's Sage):

conda install -c conda-forge "sagemath-mcp>=0.9.4" sage
sagemath-mcp

The version floor matters: earlier releases cannot start a worker on conda-forge's sage, and without it the solver may pick one. That sage is 10.9, one release behind the 10.10 the security artifacts are generated from, so names 10.10 added are unavailable and nothing outside the allowlist is admitted.

A Sage runtime without the 3 GB image (passagemath, optional):

pip install "sagemath-mcp[passagemath]"    # ~1 GB, no Docker, no local Sage build
sagemath-mcp

A pip-installable, modularized fork of SageMath. from sage.all import * and the worker run unmodified; the server detects the runtime at import and loads the matching security artifacts, so the deny-by-default policy is equivalent on both. It is pinned exactly (passagemath-standard==10.8.12) and exercised by its own CI lane — the whole suite plus the doctest-corpus sweep against the pin — so a pin bump is verified end to end before it ships (docs/passagemath_evaluation.md). It is the optional runtime; the monolithic image stays primary, and for untrusted or multi-tenant use run the container regardless of runtime — a pip install has your user's privileges, the container adds OS-level isolation.

On arm64 (Apple silicon, Graviton): the passagemath image. The monolithic image above is published for linux/amd64 only, so on arm64 it runs under emulation. The same server on the passagemath runtime ships as a native linux/amd64 + linux/arm64 image, with the same hardening flags, UID and security policy:

docker run --rm \
  --read-only --tmpfs /tmp:rw,size=512m --tmpfs /home/sage/.sage:rw,size=256m \
  --cap-drop ALL --security-opt no-new-privileges --pids-limit 256 --memory 4g \
  -p 127.0.0.1:8314:8314 \
  ghcr.io/xbp-europe/sagemath-mcp:latest-passagemath

Every release tag has a -passagemath twin (vX.Y.Z-passagemath), each architecture is smoke-tested natively before it is published, and the index is signed and attested like the primary image. It is the optional image on amd64, where the monolithic one stays primary.

One click in a desktop MCP host (Claude Desktop and friends): download sagemath-mcp-<version>.mcpb from the latest release and open it. The bundle is about 27 KB; your host installs the server and a Sage runtime with uv on first launch, which is roughly a 1 GB download once. macOS and Linux — native Windows is excluded because passagemath's Windows support is partial. A bundle is a local install with your own privileges; for untrusted or multi-tenant use run the container instead.

Source install, Docker Compose, and the Kubernetes Helm chart are in USAGE.md.

Connect an MCP client

One command or one click each. Every route below installs a Sage runtime alongside the server, so there is nothing else to set up.

Claude Desktop — download sagemath-mcp-<version>.mcpb from the latest release and open it. One click, no config file.

Claude Code — as a plugin from this repository's marketplace, inside a session:

/plugin install sagemath-mcp --marketplace XBP-Europe/sagemath-mcp

or as a plain MCP server:

claude mcp add sagemath -- uvx --from "sagemath-mcp[passagemath]" sagemath-mcp

Gemini CLI — the repository is itself an extension:

gemini extensions install https://github.com/XBP-Europe/sagemath-mcp

Codex CLI

codex mcp add sagemath -- uvx --from "sagemath-mcp[passagemath]" sagemath-mcp

Then ask for some mathematics — the examples below are a good start.

Three things worth knowing. These need uv on your PATH, and the first launch downloads about 1 GB of Sage wheels, cached afterwards. Already have sage? Drop [passagemath] from the specification and it will use yours. Installed from conda-forge? Register the environment's own executable, for example claude mcp add sagemath -- ~/miniforge3/envs/sage/bin/sagemath-mcp; it finds the sage beside it without the environment being activated. And an install of any of these runs with your own privileges; for untrusted or shared use, run the container and point the client at it.

Pinning a release, HTTP transport and the full client reference are in USAGE.md.

Try it

Prompts a client can run once the server is connected:

• Damped harmonic oscillator — "Solve x'' + 2·x' + 5·x = 0 with x(0)=1, x'(0)=0, then verify the solution satisfies the ODE."
• General relativity — "On the hyperbolic upper half-plane with metric (dx² + dy²)/y², compute the Ricci scalar and confirm it is a constant negative curvature."
• Coupled two-tank system — "Solve the linear ODE system for two mixing tanks, then take the long-term limit of each concentration."

Each builds an object once and explores it across calls — the case for evaluate_sage and its persistent session.

The 40 tools

The math tools use SageMath as the backend; parameters and examples for the most-used ones are in USAGE.md.

Plus HTTP /health and /ready endpoints and 3 MCP resources (session snapshots, monitoring metrics, doc links). Prefer interrupt_sage_session over cancel_sage_session — it stops a computation while keeping the session's variables.

How it works

┌─────────────────────────────────────────────────────────────┐
│  MCP Client (Claude Desktop, Gemini CLI, Codex CLI, ...)    │
└───────────────────────────┬─────────────────────────────────┘
                            │  MCP protocol (stdio or HTTP)
                            ▼
┌─────────────────────────────────────────────────────────────┐
│  app.py + tools/ --- FastMCP 4.x                            │
│  ┌─────────────┐  ┌──────────────┐                          │
│  │ 40 MCP Tools│  │ 3 Resources  │   manager.py routes each │
│  └─────────────┘  └──────────────┘   client to its worker   │
└───────────────────────────┬─────────────────────────────────┘
                            ▼   one subprocess per session
┌─────────────────────────────────────────────────────────────┐
│  _sage_worker.py --- allowlist.py + security.py             │
│  AST validation, then exec() in a persistent namespace      │
│  (vars, functions and classes survive across calls)         │
└─────────────────────────────────────────────────────────────┘

Request flow: MCP client → a tool in tools/ → SageSessionManager.get() → SageSession.evaluate() → JSON request to the _sage_worker.py subprocess → AST validation → exec() in the persistent namespace → JSON response.

• Process isolation — each session runs Sage in its own subprocess; a crash or timeout in one cannot affect another.
• Stateful sessions — variables, functions, and assumptions persist across calls, enabling multi-step workflows.
• Deny-by-default — a name is refused unless the generated allowlist offers it or the caller's own code bound it. A helper a future SageMath adds is refused until someone reviews it, rather than reachable the day it lands.

Security

The AST validator is defence in depth against accidents and casual misuse — it is not a boundary against determined adversarial code. The container is the security boundary. The server has no authentication, so every default is loopback: --host defaults to 127.0.0.1, the default transport is stdio, and Compose / Helm keep the endpoint off the network. Put something that authenticates in front of it before exposing it.

What the policy enforces: an allowlist (caller code may read only a name the server offers or the caller itself bound); no imports by default; eval / exec / compile and runtime string evaluation blocked; dunder access blocked; the external CAS interfaces and every file / network / persistence primitive removed from the namespace, by provenance rather than by name; and the sage package tree closed to caller code, so a helper is reached by its own name or not at all. The container adds a read-only root, dropped capabilities, no-new-privileges, the default seccomp profile, and memory ceilings. Compose adds a fork (PID) ceiling; Kubernetes has no per-pod equivalent in the pod spec, so on the chart that is the node's podPidsLimit rather than something the chart can set.

Full threat model and the complete blocked / allowed tables: SECURITY.md and USAGE.md § Security model.

Docs & more

• USAGE.md — the full manual: every tool, how code is interpreted, deployment, and the security model.
• docs/mcp_quickstart.md — a first-session walk-through.
• CHANGELOG.md · ROADMAP.md · CONTRIBUTING.md · SECURITY.md · SUPPORT.md — what "supported" means, release cadence, where to ask. PRIVACY.md — what the server keeps (nothing leaves the machine it runs on).

Requirements

Python 3.12+ and a SageMath runtime (the container image bundles SageMath 10.10; otherwise sage on PATH, or the [passagemath] extra). Built on FastMCP 4.x.

Contributing

Issues and pull requests welcome — see CONTRIBUTING.md. Run make lint, make typecheck and make test before pushing (git config core.hooksPath .githooks wires the pre-push check). Roadmap and open work: ROADMAP.md. Questions go to GitHub Discussions; SUPPORT.md says what to expect.

Citing

If this server is part of published work, cite it via CITATION.cff — GitHub renders it as Cite this repository in the sidebar, with APA and BibTeX. Cite SageMath itself as well; it does the mathematics.

License

MIT — see LICENSE. SageMath itself is GPL-2.0-or-later and is used as a separate runtime; no SageMath source is redistributed in this repository.