Skill Atlas which agent skills are worth usinghosted preview

For agents

API address. Hosted: https://0xquez-skill-atlas-api.hf.space (MCP at /mcp). Locally the same service is at http://127.0.0.1:8765. The hosted Space may sleep when idle; the first request after that can take a minute. It is rate-limited to 60 requests per minute per IP.

Agents can ask Skill Atlas about a skill before they install it or follow it. Three steps:

  1. Search for what the skill should do.
  2. Read the verdict. use means go ahead. caution means read the reasons first. ignore means don't use this one.
  3. If it's a copy, use the original. The canonical_id field points to it.

Each answer also says why (reasons), lists risky patterns (risk_flags), says whether it's a slop-skill, and links to GitHub. Security tools carry category: "white_hat_security" and a white_hat_note: they do dangerous things on purpose, so use them only for security work you're allowed to do. The tag never changes the verdict.

Search rules

Every word you type must appear somewhere in the skill's name, its GitHub owner/repo, or the start of its description. Word order doesn't matter. Results show exact owner/repo matches first, then skills with all your words in the name, then use before caution before ignore, then originals before copies, then more GitHub stars. The website search uses the same rules.

MCP tools

Tool What you send What you get
search_skills query, optional verdict (use, caution, ignore), optional slop (true/false), optional white_hat (include, exclude, only; default include), limit (default 10, max 100) a ranked list of skills
get_skill_verdict skill_id one skill: verdict, reasons, scores, link to the original
find_canonical skill_id the original version of this skill, or a note that no safe original exists

Connect an MCP client

{
  "mcpServers": {
    "skill-atlas": { "url": "https://0xquez-skill-atlas-api.hf.space/mcp" }
  }
}
# pip install "mcp>=2.3"
import asyncio
from mcp import Client

async def main():
    async with Client("https://0xquez-skill-atlas-api.hf.space/mcp") as c:
        res = await c.call_tool("search_skills", {"query": "terraform module", "verdict": "use", "limit": 3})
        print(res.content[0].text)

asyncio.run(main())

Web API

Method Path Notes
GET /skills/search?q=&verdict=&slop=&white_hat=&limit= same rules as above
GET /skills/{skill_id}/verdict one skill
GET /skills/{skill_id}/canonical the original version
GET /stats headline numbers
GET /healthz {"ok": true}
curl -s "https://0xquez-skill-atlas-api.hf.space/skills/search?q=react+native+testing&verdict=use&limit=3" | jq '.results[] | {skill_id, name, verdict, github_url}'

What the reasons mean

Reason Effect Meaning
known_malicious_source ignore The text comes from a collection of deliberately harmful test skills, or copies its harmful part.
near_copy_of:<id> ignore A copy. Use skill <id> instead.
renamed_copy ignore A copy published under a different name.
slop_skill ignore One of many template look-alikes from one account.
risk_high ignore Risk score 40 or more, for example code that reads passwords or keys and sends data out.
quality_low ignore Too thin or broken to be useful.
junk:* ignore Empty, a redirect, or too short.
risk_medium, risk_flag:<flag> caution Some risky patterns. Read before running.
quality_medium, invalid_frontmatter caution Usable but rough.
translated_copy_of:<id> none A translation of skill <id>. For information only.

A simple policy for agents

1. search_skills(query, verdict="use") and take the top result.
2. If nothing fits, search without the verdict filter, then call find_canonical on a hit.
3. Never install a skill marked "ignore". For "caution", show the reasons to the user first.