# SkillProof MCP

Ask whether a Claude Code skill actually works — **before** you install it.

GitHub has tens of thousands of `SKILL.md` files. Almost none have been run by anyone but their author.
[SkillProof](https://skillproof.dev) installs them from their repo, triggers them, and runs them on a real
task against a no-skill baseline. This MCP server puts those verdicts in your agent's hands.

```
> is there a tested skill for converting markdown to Confluence?

## Confluence — DIDN'T PASS — scored below the no-skill baseline
What our test found: Ran the bundled convert_markdown_to_wiki.py on a real sample doc: it silently
turns **bold** text into wiki _italic_ (a genuine regex bug), and leaves standard GitHub-style tables
completely unconverted, despite SKILL.md listing "tables" among the elements it handles.
```

That is the whole point. A directory that only lists winners tells you nothing.

## Tools

| Tool | What it answers |
| --- | --- |
| `find_skill` | "Is there a tested skill for *X*?" — ranked matches with verdict, score, test notes, install command |
| `check_skill` | "Someone recommended *X* — is it any good?" — the verdict for one skill by name, slug, or repo |

Every answer carries one of three verdicts:

- **pass** — installed, triggered, and beat the no-skill baseline on a real task.
- **setup** — works, but needs a manual step first (the notes say which).
- **didn't pass** — scored below the no-skill baseline: it either couldn't run, or left you worse off
  than not installing it.

If nothing has been tested for your job, the server says so instead of guessing. "Not tested" is a real answer.

## Install

Claude Code:

```bash
claude mcp add skillproof -- npx -y skillproof-mcp
```

Or add it to your MCP config by hand:

```json
{
  "mcpServers": {
    "skillproof": {
      "command": "npx",
      "args": ["-y", "skillproof-mcp"]
    }
  }
}
```

Works in any MCP client (Claude Code, Claude Desktop, Cursor, Windsurf, Zed). Node 18+. No API key, no account.

## Where the data comes from

The server reads the live catalog at [`skillproof.dev/api/skills.json`](https://skillproof.dev/api/skills.json)
and caches it for 15 minutes. Nothing is bundled, so verdicts are never stale. The scoring rubric —
install /5, triggering /5, output-vs-baseline /10, docs /5 — is published at
[skillproof.dev/methodology](https://skillproof.dev/methodology).

Catalog data is CC BY 4.0: use it, cite `skillproof.dev`.

## License

MIT