io.github.Skillproofdev/skillproof
Check whether a Claude Code skill was tested and whether it works before installing it.
Open source Open in the app JSON README (API)
About
Check whether a Claude Code skill was tested and whether it works before installing it.
Details
- Kind
- MCP servers
- Topic
- No topic detected
- Publisher
- skillproofdev
- Origin
- official
- Category
- ferramentas
- Transport
- local
- Version
- 1.0.1
- Last push
- 2026-07-14T21:50:27Z
- Repository state
- ativo
- Language
- JavaScript
- License
- MIT
- Added
- 2026-08-29 03:02:15
- Updated
- 2026-08-29 03:02:15
- Origin id
io.github.Skillproofdev/skillproof
README
# 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