Scoring MCP servers 0–100 (A–F): the 5 hygiene categories — and where would YOU set the CI gate?

[manja316]

Third thread, different from the methodology one (#1) and the manual 60-second checklist (#2). This one is about the number the tool spits out — the 0–100 hygiene score / A–F grade — and the thing I think is actually underused: wiring that grade into CI as a pass/fail gate.

What the score is made of

audit_scan (or audit_quick_scan for just the headline) connects to a running MCP server, enumerates every tool/resource/prompt, and scores hygiene across 5 categories, 100 points total:

Category	Points	What it rewards
Documentation	25	Every tool has a real description, params explained
Schema rigor	25	Typed params, enums/bounds, no free-form string catch-alls
Injection safety	25	Tool descriptions can't be turned into instructions to the model
Scope signals	15	The server signals what each tool can reach (file/network/shell)
Metadata	10	Versioning, server name/description present

Grade bands: A ≥ 90, B ≥ 80, C ≥ 65, D ≥ 50, F < 50.

The part most people miss: it's already a CI gate

The CLI exits 0 only when the grade is A or B, non-zero otherwise. So this is a one-liner in a pipeline:

- run: |
    pip install mcp-security-audit
    mcp-audit scan ./your-server   # exits 1 if grade drops below B

A server that was an A and slips to a C because someone added an untyped shell-reaching tool with a vague description will now fail the build — before it ships to anyone's Claude/Cursor config.

The open question

Where would you set the gate? A/B (the current default) is strict — a single undocumented tool can knock a small server out of B. Is C "good enough to ship" for an internal-only server? Should injection-safety be a hard fail regardless of total score (i.e. you can't pass with 25/25 docs but 0/25 injection)?

I lean toward making injection_safety a hard floor independent of the total — a server can be beautifully documented and still be an exfil path. Curious whether people running MCP servers in prod agree, or think that's too aggressive for a first gate.

---

Free + OSS: pip install mcp-security-audit · github.com/LuciferForge/mcp-security-audit (a star helps if it's useful). Methodology in #1, manual checklist in #2.