mcpseal

Source, issues and docs: https://github.com/confuseddude/mcpseal

An MCP tool-integrity CLI: pins the hash of every approved MCP tool definition and blocks execution the instant a tool's description or input schema drifts — the pattern known as a "rug pull," where a server changes a tool's behavior after you've already trusted it. Free, local, zero-infra.

What it does

1. mcpseal init launches each MCP server your client is configured to use, records every tool's name/description/input schema as a cryptographic hash, and writes a .mcp-lock.json you commit to your repo (like package-lock.json).
2. mcpseal install rewrites your client's config so it launches your servers through mcpseal proxy instead of directly.
3. From then on, every time a tool definition is served to your client, mcpseal re-hashes it and compares against the pinned hash. A match forwards normally. A mismatch gets blocked before your client ever sees it, with the exact old-vs-new description shown.

Privacy

Nothing leaves your machine, ever, unless you explicitly run mcpseal login. No telemetry, no network calls, no account required. login is entirely optional and only relevant if your organization runs the paid Control Plane — everything above is the complete product for a solo developer.

Quickstart

uvx mcpseal init
uvx mcpseal install

or, if you prefer a persistent install:

pip install mcpseal
mcpseal init
mcpseal install

Commands

Full source, architecture docs, and the Node.js twin of this CLI are in the same monorepo this package is built from.

Verifying what you installed

Every release is published from GitHub Actions with no stored credentials, and carries build provenance you can check yourself:

mcpseal --version
npm audit signatures            # npm: verifies the SLSA provenance attestation

PyPI attestations (PEP 740) are shown on the project page and name the exact repo, workflow and commit that produced the artifact.

Reporting a vulnerability

See SECURITY.md.