The CLI

agentcheck runs the same checks as this site, from your machine. No account, no signup, no dependencies beyond Node 18.

npx agentcheck yoursite.com

It is open source, MIT licensed.

The grades match

The CLI and the hosted product share their scoring deliberately. A tool that grades your site B while the website grades it D would make both untrustworthy, and the disagreement is what people would remember.

Options

--spec <name>      Check one spec: ucp or webmcp
--json             Machine-readable output
--markdown         Markdown, for a pull request comment
--fail-on <level>  Exit 1 on: fail (default), warn, never
--min-grade <A-F>  Exit 1 if any checked spec grades below this
--timeout <ms>     Per-request timeout

In CI

This is where it earns its keep. A check in your pipeline catches a regression the day it ships, rather than whenever somebody next thinks to look:

- uses: orbilo/agentcheck@v1
  with:
    url: https://example.com
    min-grade: B

The action writes the report to your job summary and exposes grade, score, failures, warnings and indeterminate to later steps.

Exit codes

An inconclusive result — a rate limit, a timeout, a 503 — exits 2, never 1. Being rate limited is not the same as being broken, and failing a build over one teaches people to stop trusting the exit code.

It works against localhost

npx agentcheck localhost:3000 --spec webmcp

Loopback counts as a secure context, exactly as it does in browsers, so you can check your work before it ships rather than after.

What the CLI cannot do

Tell you it still passes tomorrow. Nothing on your machine is watching between runs, so a check that passes today says nothing about the morning a platform update strips a header. That is what adding a site is for.