SkillsMP is a public directory of agent skills scraped from GitHub. It indexes around 1.3 million skills compatible with Claude Code, Codex CLI, and ChatGPT, lets you filter by occupation, and installs to ~/.claude/skills/ with a copy-paste. Independent community project. Not affiliated with Anthropic.
Site: skillsmp.com
The problem nobody talks about
Finding good agent skills today means trawling GitHub. Search returns repo names, not capabilities. You read READMEs, guess if the skill works for your stack, copy the folder, and hope. There's no central index, no quality signal, no occupation filter. The ecosystem is already bigger than any one person can scan.
Why people are split
The love camp says SkillsMP solved discovery. 1.3M skills indexed, filterable by job (867 SOC categories), sorted by GitHub stars. You can find a skill for a specific role in under a minute.
The hate camp says volume is the problem, not the solution. Most of those 1.3M are auto-scraped low-effort skills. A "minimum 2 stars" quality filter doesn't catch much. And skills are code that runs on your machine. A directory that makes one-click install feel safe trains people to skip reading the source.
Both sides are right. The site is useful if you treat it as a search engine. Dangerous if you treat it as an app store.
How to use it without getting burned
- Search by what you want to do. The occupation filter is noisy.
- Sort by stars, not recency. Recency promotes spam.
- Click through to the GitHub repo before installing. Read the
SKILL.mdand anyscripts/*files. - If a skill runs shell commands or hits a network endpoint, understand what it touches before adding it to your config.
Honest limitations
- Quality is a coin flip. The 2-star minimum is a low bar.
- The "one-click run via Manus" path bypasses local review entirely. Skip it.
- No version pinning. The skill you install today can change tomorrow if the repo does.
- Community project with no SLA. Don't rely on it for anything production.