| name | mcp-audit |
| description | Audit an MCP server for conformance — lint its tool schemas and behaviorally fuzz it (call tools with malformed input), then score it 0–100 (grade A–F) with the exact findings and fixes. Use when the user wants to check, test, validate, score, benchmark, or debug an MCP server, its tools or tool schemas, or figure out why an MCP server/tool behaves badly with an agent. |
Audit an MCP server (MCProbe)
MCProbe scores an MCP server's conformance — how reliably an AI agent can
actually use it — across schema quality and behavioral robustness. Reach for it
whenever the user wants to evaluate, test, compare, or debug an MCP server.
Pick the path that's available
A. MCProbe is connected as an MCP server (you have probe_* tools).
Run the four core tools in order:
probe_connect — dial the target: an HTTPS URL, or a stdio command.probe_report — the one-shot audit: lints the schemas, fuzzes behavior, and
returns the full report + 0–100 score. (Use probe_lint / probe_fuzz
individually if you want just one phase.)probe_disconnect when finished.
B. The mcprobe CLI is available (no MCP tools connected).
npx mcprobe audit https://example.com/mcp --fuzz
npx mcprobe audit --stdio "npx @acme/my-mcp-server" --fuzz
npx mcprobe audit https://example.com/mcp --fuzz --json
Omit --fuzz for a fast static (schema-only) audit. Add --json to parse the
result programmatically.
What --fuzz does — and the safety rule
Fuzzing calls each tool with malformed input to test error handling and
liveness. This is what catches the dangerous problems a linter can't: a tool
that silently accepts garbage (so the agent trusts a wrong result) or one that
crashes the session. It is dry-run by default — tools annotated
destructiveHint: true are skipped unless you pass --fuzz-destructive, so a
normal run is safe even on servers you don't own. Only fuzz servers you're
allowed to test.
How to read the result
- Overall 0–100 + grade A–F (A ≥ 90, B ≥ 75, C ≥ 60, D ≥ 40, F < 40).
- Four dimensions, each 0–10: schema quality, error handling, liveness, and
metadata/annotations.
- Critical-issues callout: the behaviors that hurt agents and users most, so
surface them first — tools that silently accept bad input, tools that
return an empty success on a valid call (a hallucinated success: the tool
reports "done" but returns/persists nothing), tools that return a success that
violates their declared
outputSchema (the payload doesn't honor the
contract the tool advertised), and protocol crashes.
- Findings carry stable dotted codes (
tool.missing_description,
param.untyped, schema.invalid, …) and fix hints. Report the worst-severity
ones with the exact tool/parameter and the recommended fix.
Optional: save the run to a dashboard (needs an mcprobe.org Pro key)
push runs the same audit and uploads it to the user's mcprobe.org history
and public gallery. The audit itself is always free and local; only the hosted
upload is the Pro tier.
How the user gets a key (one-time):
- Go to https://www.mcprobe.org and sign up (email + password, then
confirm via the email link).
- Go Pro — $9.90 once, lifetime. Click "Go Pro" / "Unlock" and pay by
crypto (card checkout is coming soon).
- Open My Profile → https://www.mcprobe.org/app/profile.
- In the "Audit a local (stdio) server" card, click Generate key. A
token like
mcp_… appears — it's shown only once, so copy it right away.
- Store it as an environment variable (keeps it out of shell history / code):
export MCPROBE_TOKEN="mcp_your_key_here"
Then upload any audit:
npx mcprobe push https://example.com/mcp --fuzz --token "$MCPROBE_TOKEN"
The CLI also reads MCPROBE_TOKEN from the environment, so --token can be
omitted once it's set.
More detail (scoring model, the 12 lint rules, the fuzz categories):
https://github.com/alitiknazoglu/mcprobe
