Know what an MCP server asks for before the pull request merges.

Install the GitHub App on a repository. When a pull request adds or changes an MCP server or a skill, it gets one comment and one check with what we know about it. The whole team sees it where they already review.

Free. Per repository or for a whole organization. It is informational by default: it does not fail a build and does not change what runs.

What the comment says

For each server or skill the pull request adds or changes, four answers and the result as our engine scores it, with a link to the full result. One comment per pull request. A later push edits it; it never posts a second one.

A server listed in several config files counts once. A hosted server with no package is shown as "hosted · no package". A package we have no result for is queued, and the comment is updated when there is one. We do not guess a result from a name.

An example of the comment, with an invented packageKillerSkills: 1 agent tool added or changed: 1 Cautionexample-notes-mcp MCP server · added in .mcp.json
What it asks for
NOTES_API_TOKEN
Where it sends it
api.example-notes.dev (its own).
What we saw
No credential files read. It reached only api.example-notes.dev.
What we couldn't test
Tool calls that need a real account.
Result: Caution, 69 for version 1.4.2, run 6 days ago. Full resultThis check is informational. It does not fail a build and does not change what runs. A result describes what one version did in the tests that ran.

What it can read, and what it keeps

ContentsReadTo read the changed files that declare agent tools: MCP config files, SKILL.md and the files beside it, and package.json. Other files in a pull request are not fetched.
Pull requestsRead and writeTo list the files a pull request changes and to post one comment, which is edited in place on later pushes.
ChecksRead and writeTo post one check run per push. It is informational unless the repository opts in to failing on a Blocked result.
MetadataReadRequired by GitHub for every app.

It stores names, never code or secret values. For a pull request it comments on, we keep the repository name, the pull request number and the names of the tools found. File contents are read to find those names and are not kept. A credential value written into a config is reported by the name it sits under; the value is never stored or posted.

Fail the check on a Blocked result

Off unless you turn it on. Add a file named .killerskills.yml to the default branch:

fail-on: blocked

It is read from the branch the pull request targets, so a pull request does not switch its own check off. Caution and "no result yet" never fail the check.

A badge for a public repository

It links to the repository's scan page. Replace OWNER and REPO:

[![MCP checked by KillerSkills](https://killerskills.ai/badge/repo/OWNER/REPO.svg)](https://killerskills.ai/scan/github/OWNER/REPO)

The badge says a scan exists. It is not a statement that the repository is safe.

What this does not do

  • It does not run anything in your repository or on your machines. It comments and posts a check.
  • A skill that lives in your repository is read, not run, and gets no label when the reading finds nothing.
  • It does not detect prompt injection. The repository's own instruction files, such as CLAUDE.md, are not reported.
  • Limits on what an agent can do on a developer's machine are a separate, optional step: the local guard. Nothing is enforced at run time unless someone installs it.