
mcpwatch
io.github.anardimontlv/mcpwatch
Check MCP servers' security grades and tool changes before your agent connects. Read-only.
Documentation
mcpwatch
Know when the tools your AI agents trust change.
MCP servers can quietly rewrite a tool's description after you approved it (a "rug pull"), and that description is read by your agent as instructions. mcpwatch checks every server in the official MCP registry daily, fingerprints the tools of every publicly reachable remote server, grades each server, and fails your CI when something you pinned drifts.
- Static and read-only. It never runs third-party server code and never calls a tool.
- Private by default. Your config never leaves your machine; the CLI only downloads public index files.
- Zero dependencies. Python 3.10+ standard library.
Live index and rug-pull radar: https://www.mcpwatch.io
Quickstart
pipx install mcpwatch-io # or: pip install mcpwatch-io
# 1. Scan a config for risky setups (leaked secrets, unpinned packages, root filesystem access, plain HTTP, ...)
mcpwatch-scan ~/.cursor/mcp.json
# 2. In your project: pin the exact tools your servers expose today...
cd your-project
mcpwatch lock .mcp.json # writes mcpwatch.lock: commit it
# 3. ...and fail when any of them changes
mcpwatch check .mcp.json # exit 1 on drift, 2 if the index can't be reached
Or paste a config into the browser scanner at https://www.mcpwatch.io/scan (runs locally in your browser). Tool fingerprints exist for remote servers we can reach without credentials; package-based servers are matched by name and version, and their tools are not fingerprinted (we never run their code).
Works with Claude Desktop, Claude Code (.mcp.json), Cursor and VS Code configs.
GitHub Action
Scan on every pull request:
- uses: actions/checkout@v4
- uses: anardimontlv/mcpwatch-action@v1
with: { config: .mcp.json, fail-on: high }
Check for drift every day (rug-pull radar for your repo):
on: { schedule: [{ cron: "0 6 * * *" }] }
jobs:
mcp-drift:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: anardimontlv/mcpwatch-action@v1
with: { mode: check, config: .mcp.json, lock: mcpwatch.lock }
What check treats as drift
| Change | Severity |
|---|---|
| A pinned tool's description or input schema changed | high |
| Endpoint stopped requiring credentials | high |
| Server disappeared from the index, or its lookup now points elsewhere | high |
| Tool added or removed; grade changed | medium |
| Endpoint now requires credentials, or tools can no longer be verified | medium |
Fingerprints older than --max-stale days (default 3): a server can't hide by blocking the indexer | medium |
Server in the config but not in the lock file (not_locked) | medium |
| Package version changed; server newly indexed; server removed from config | low (reported, does not fail) |
Scanner rules
| Rule | Finds |
|---|---|
| MW001 | Secrets in config: env, headers, args, URLs, connection strings |
| MW002 | Unpinned packages (npx pkg, uvx pkg@latest) |
| MW004 | Remote server over plain HTTP |
| MW005 | Remote server without auth |
| MW006 | Filesystem root or whole home directory granted |
| MW007 | Privileged containers, host mounts, Docker socket |
| MW008 | Inline shell launch, curl | sh |
| MW009 | Flags that disable safety controls |
| MW101 | Tool poisoning: instructions aimed at the model in a tool description |
| MW102 | Tools that execute commands or delete data |
| MW103 | Invisible Unicode in tool text |
Full rule reference: docs/rules.md. Methodology: docs/how-it-works.md. MCP server for agents: docs/mcp-server.md.
Grades are automated signals, not audits. Maintainers can dispute a finding from the server's page.
Plans
The scanner, the Action and the public index are free. Paid plans add watchlist alerts (email/Slack), full change history and an org-wide view. See https://www.mcpwatch.io/pricing.
License
Apache-2.0. See LICENSE and NOTICE. The hosted index, radar and paid plans are a commercial service.
mcpwatch-iopip install mcpwatch-io