Skills are everywhere.
Proof isn't.

We test AI agent skills for safety and real usefulness, and show the results.

The state of agent skills

What independent audits found in 2026.

Dial: Weak36.8%of audited skills had a security issueSnyk ToxicSkills · 3,984 skills · Feb 2026
Dial: Danger13.4%had a critical one: prompt injection, malicious codeSnyk ToxicSkills · Feb 2026
Dial: Excellent+16.2 ptspass-rate gain when an agent uses a curated skillSkillsBench · 7,308 runs · Feb 2026
Dial: Weak16 / 84tasks got worse after adding a skillSkillsBench · Feb 2026

Catalogs explode

Skills listed per marketplace. Counts measure how much gets listed, not how good it is.

1k10k100k1M10MJanFebMarAprMayJunJulAugSepOctNovSkillsMPskills.shClawHubLobeHubMCP MarketTessllatest, undated
SkillsMP: 4.1M, own count (mostly forks)skills.sh: 69k early 2026 → ~670k JunClawHub: 5.7k → 44k in 2 monthsLobeHub: 335kMCP Market: 73kTessl: 3k
Market report 17 Nov 2026: Vercel, Similarweb via third parties, Koi Security, Conscia, skillsmp.com, LobeHub, MCP Market, Tessl. skills.sh early-2026 point is approximate.

Does a skill help?

Pass-rate change, percentage points.

SkillsBench, arXiv 2602.12670 · Feb 2026

Is it safe?

3,984 audited skills.

Critical · 13.4%Other issue · 23.4%None found · 63.2%
Snyk ToxicSkills · Feb 2026

Two scores. Never blended.

How we score
Dial: Good

Critic Score 82 / 100

Good: our reproducible tests.

4.6

Crowd Score 4.6 / 5

128 reviews from people who use it

Scores agree

Test results

Starts at the right time88%
Gets the job done76%
Safe to install100%
Easy to follow82%
Allowed to use92%

Example values for illustration.

Built a skill?

Submit a public repository. A person checks every submission before it is listed.

Submit a skill