Launching today
Agent Skills
Find skills for Claude Code, Cursor, Copilot & more
120 followers
Find skills for Claude Code, Cursor, Copilot & more
120 followers
Largest cross-platform directory of AI agent skills. 100K+ skills, 30+ platforms, security audits on every listing. One search, every platform.
Free
Launch Team / Built With
Melies
Hey everyone! Maker here.
I've been using Claude Code skills for months and they're genuinely incredible.
For me, they're like the scene in The Matrix where Neo gets Kung Fu uploaded directly into his brain.
A single SKILL.md file can completely transform how your agent works and make it learn about SEO, how to write cold emails or even how to do accounting in France.
However finding good ones? Total mess. Scattered across thousands of GitHub repos with no way to search or compare. Then the ClawHub malware incident happened. 20% of submitted skills were malicious. Prompt injection, credential theft, obfuscated code.
So I built agentskill.sh that currently indexes 100k+ skills for Claude Code, Cursor, Codex, Windsurf and more; I focused on two things:
Security: Every skill gets scanned across 12 threat categories so you know what you're installing before you install it. You can check the details here: Security Dashboard
Discovery: You can search skills by categories using many criteria, review them, and more.
The fastest way to try it is the /learn command. Once installed you (or your agent) can search and install skills directly using:
Using /learn to find skills has another big advantage: it lets your agent learn by itself.
When your agent hits a problem it doesn't know how to solve, it can search for and install the right skill on its own. No manual hunting, no copy pasting from GitHub. Your agent just gets smarter as it works.
What makes /learn special:
Two layer security. Every skill on agentskill.sh is scanned server side for 12 threat categories (command injection, data exfiltration, credential harvesting...). Then /learn performs a second client side verification before installing. You get both centralized scanning and local confirmation.
Feedback loop. Your agent auto rates skills after using them, so the best ones surface and broken ones get flagged by the community. Your agent contributes to, and benefits from, collective quality signals.
No context switch. Search 100k+ skills mid conversation, install what you need, and keep working.
Would love to hear what skills you're using and have your feedback on this!
@owner @romainsimon Really interesting concept.
One thing Iâve been noticing lately is that as more AI agents appear, discovery is becoming almost as important as the models themselves. Tools that organize the ecosystem might end up being incredibly valuable.
Curious what youâre seeing so far â are most users coming from developers building agents, or people experimenting with AI workflows?
Melies
@hpsimulator exactly, discovery is the bottleneck right now.. It's a mix of both for the moment but i think a growing number of non-technical users now use this (via Cowork or even Claude code directly)
The security scanning is the real differentiator here. 20% malicious rate on submitted skills is wild but not surprising given how the ecosystem exploded. Discovery is the other half that matters. Right now finding good skills is like searching GitHub in 2010. Having a curated, searchable index with security scores changes the game for teams that want to adopt agent skills without rolling the dice on supply chain attacks.
Melies
@devon__kelley Yes, the security piece is non-negotiable, especially since vulnerabilities can be more subtle when executing text instruction instead of code directly. More about it here: https://agentskill.sh/security
@romainsimon That's a great point about text instructions being harder to audit than code. A malicious SKILL.md can look totally benign to a human reviewer but still exploit the agent's context. Will check out the security page, curious how you handle obfuscated intent specifically.
Melies
@devon__kelley I don't have a perfect solution but currently here's how it works :
All skills are scanned for known patterns by agentskill.sh when added on the skill registry
the /learn install command forces the agent to run automated scanner locally if you have one installed, or read the entire skill before install by following this SECURITY.md embedded in the install skill. And it can report back to the registry we can take down any skill that might have passed the static scan from 1.
Melies
@joel_farthing Spot on! That's exactly how it works. /learn with no arguments scans your project (package.json, file extensions, config files, even your current git branch) and recommends skills based on what it finds.
So yes, in a Next.js + Prisma project it would suggest relevant skills for that stack. And if you're on a feat/stripe-checkout branch, it picks up on that too.
This also means your agent can self-improve. It detects what it's working on, finds the right skills, and levels up on its own. No manual searching needed.
Thanks for testing it out!
@romainsimon Could the self-improvement be somewhat automated with a user rule, like "During project planning, query agentskill.sh and install relevant expertise via learn /"?
Melies
@joel_farthing Yes definitely. For example now Claude Code has a /loop feature.
I haven't tried this but let me know if you do!
I spend a serious amount of time hunting down and wiring up Claude Code skills for our accounting automation stack â discoverability is the real friction. Most skill repos are scattered or undocumented. Centralizing this is the right call. Curious: do you curate for quality before listing, and is there a way to pin to a specific skill version so a production workflow doesn't break when a skill gets updated?
Melies
@slavaakulov Those are two excellent points.
Curating the best skills
Currently all skills are added but they are ranked depending on their security score and usage (github stars, etc).
However I built in an auto-review mechanism that will hopefully help surface the best skills : if you use the /learn command to install skills, it allows agents to report back if a skill has been useful once used for the first time.
Skill Versionning
You cannot pin down a specific version, I need to implement that. However, it currently tracks which version of the skill is installed via a content sha. When the skill is used, it should ask you if you want to update to the latest version if you don't have the latest one.
Indie.Deals
The URL ends up in a 404, you may want fix the URL as it contains "@"
https://agentskill.sh/@?ref=producthunt
Melies
@adithya Thanks! it's fixed. The url with the ref was weirdly detected as a skill url
AutonomyAI
If this helps make skill building and selection even 10% easier it's worth its weight in gold!
Happy to support - best of luck!
Melies
@lev_kerzhner Thanks a lot ! I just added more install methods so now you can:
use the /learn skill to install other skills (best method since it handles security, review, updates, ...)
Copy install prompt (no preinstall needed, it should just work)
Download the zip (useful for people using Claude Cowork for example)
Needle
Cool graphics!
Melies
@jan_heimes Thanks :)