The MCP ecosystem has 15,000+ servers — and roughly 1 in 6 has a verified problem: dead repos, deprecated packages, stale code, leaked secrets. mcp-grade is the fast, opinionated linter + Lighthouse you run before you wire a server into your agent.

✨ Why

• One command, instant grade. No config, no signup, <5 seconds.
• Security first. Catches hardcoded secrets, committed .env files, and eval/exec — weighted so a leaky server can never score well.
• Prove it actually runs. --live launches the server and verifies the real MCP handshake.
• Shareable report cards. Export a Markdown card + shields badge for your README or PRs.
• CI-native. A GitHub Action + --fail-under make it a quality gate that fails the build.
• Ecosystem leaderboard. Rank popular servers and publish a public scoreboard.

📦 Install

pipx install mcp-grade                 # core CLI
pipx install "mcp-grade[live]"         # + live protocol probing
pip install "mcp-grade[web]"           # + the leaderboard web dashboard

🚀 Quickstart

mcp-grade ./path/to/server                       # grade a local repo
mcp-grade ./path/to/server --report --badge      # + Markdown card + README badge
mcp-grade https://github.com/someone/server      # clone & scan a remote repo
mcp-grade ./path/to/server --json report.json    # machine-readable output
mcp-grade . --fail-under B                       # CI gate

🔬 Live protocol probing

Static analysis tells you a server looks right; --live proves it runs and answers. It launches the server over stdio and performs a real MCP handshake (initialize → list_tools):

pipx install "mcp-grade[live]"
mcp-grade ./path/to/server --live
# optionally control the launch command:
mcp-grade ./path/to/server --live --command "python server.py"

Live results (live_handshake, live_lists_tools) are folded into the Protocol score. It's off by default so the core scan stays fast and offline.

🏆 Leaderboard

Rank a set of servers (or the whole ecosystem) and publish a scoreboard:

mcp-grade-leaderboard ./serverA ./serverB --html leaderboard.html --json lb.json
mcp-grade-leaderboard --from-file leaderboard/repos.txt --md leaderboard.md
mcp-grade-leaderboard            # no args → scan the curated list of popular servers

Serve it as a live dashboard (deploy to Fly.io / Render / Railway):

pip install "mcp-grade[web]"
MCP_GRADE_TARGETS=./examples/good_server,./examples/bad_server \
    uvicorn mcp_grade.webapp:app --reload
# → http://127.0.0.1:8000  (HTML)  and  /api/leaderboard  (JSON)

🤖 GitHub Action

Run mcp-grade automatically on every pull request — and fail the build if a server drops below your bar.

# .github/workflows/mcp-grade.yml
name: mcp-grade
on: [push, pull_request]
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: lipon101/mcp-grade@v1
        id: scan
        with:
          fail-under: B
      - name: Show grade
        if: always()
        run: echo "Grade ${{ steps.scan.outputs.grade }} (${{ steps.scan.outputs.score }}/100)"

Every repo that adds this Action becomes a dependent — visible under Insights → Dependents. New servers can start from mcp-server-template, which has this gate pre-wired.

🧪 What it checks

Grades: A ≥ 90 · B ≥ 80 · C ≥ 65 · D ≥ 50 · F < 50.

🤝 Contributing

Contributions are welcome! See CONTRIBUTING.md. Add a check in mcp_grade/checks/, add a test, open a PR. Ready to ship? See LAUNCH.md.

📄 License

MIT © 2026 Md. Torikul Islam Lipon.