io.github.liminalpepe/agent-guard-mcp

agent-guard-mcp

One line: Verify-before-act safety for AI agents — check packages, lockfiles, skill manifests, and CI workflows before install or merge. Zero API keys. Zero config.

Scores Cursor/Claude skill manifests + GitHub Actions workflows — not just npm dependencies.

When you…	Call
npm install foo / agent suggests a new dep	check_package
Review an LLM-generated lockfile	verify_lockfile
Install a Cursor skill or MCP plugin	score_manifest
Merge a PR touching .github/workflows/*	check_workflow

Call these before installing a dependency, merging a CI change, or installing a third-party skill/plugin — so an agent refuses hallucinated packages, typosquats, poisoned manifests, and compromised CI actions.

Exposes four MCP tools (also available as a plain HTTP service):

Tool	Use before…	Returns
check_package	adding a single npm/PyPI dependency	OK / SUSPICIOUS / DANGER + typosquat/slop flags
verify_lockfile	running npm/pip/yarn install	scans every direct + transitive dep in a lockfile
score_manifest	installing a Cursor/Claude skill or MCP/Smithery plugin	poison-signature + scope-overreach + drift score 0–100
check_workflow	merging a PR that touches CI	flags mutable action pins, compromised actions, curl|bash, pwn-requests, secret exposure

Why

AI agents routinely hallucinate package names (slopsquatting), trust LLM-generated lockfiles, install unvetted skill packs, and pin CI actions to mutable tags. Each is a live supply-chain vector. This server gives an agent a cheap "is this safe?" call before it acts.

Install (MCP client)

{
  "mcpServers": {
    "agent-guard": { "command": "npx", "args": ["-y", "@liminallablibs/agent-guard-mcp@0.2.1"] }
  }
}

Agent rule (paste into Cursor Rules / CLAUDE.md)

Before installing any npm/PyPI package, running install from a lockfile, installing a third-party skill/MCP plugin, or merging CI workflow changes, call the agent-guard MCP tools (check_package, verify_lockfile, score_manifest, check_workflow). If verdict is SUSPICIOUS, REVIEW, or BLOCK/DANGER, stop and show findings to the user — do not proceed.

Quick demo

npx -y @liminallablibs/agent-guard-mcp@0.2.1   # MCP clients spawn this via npx
# HTTP mirror:
npm run http && curl "http://localhost:8402/check?name=reactt&ecosystem=npm"
# → DANGER — typosquat of "react"

Or run directly:

npm install
node src/mcp-server.mjs        # stdio MCP server
npm run http                   # optional HTTP mirror on :8402

HTTP endpoints (mirror of the MCP tools)

• GET /check?name=<pkg>&ecosystem=<npm|pypi>
• POST /check-lockfile {lockfile_content, format} — format ∈ package-lock.json | yarn.lock | pnpm-lock.yaml | poetry.lock | requirements.txt
• POST /score-manifest {manifest_type, manifest_content}
• POST /check-workflow {workflow_content}

License

MIT.