skills.sh vs SkillzBench: Which Agent Skills Platform Should You Trust?
skills.sh is Vercel's free, open directory and MIT-licensed CLI for discovering and installing agent skills across 20 coding agents, with 83,627 listed skills and 8M+ installs — but it explicitly cannot guarantee the quality or security of community-submitted skills. SkillzBench is AIQ Labs' independent testing lab, currently in pre-launch, that is being built to verify, security-scan, and benchmark skills so you can see which ones actually help. For free discovery today, skills.sh wins; for evidence on whether a skill helps or is risky, SkillzBench is the one building that answer.
TL;DR
- skills.sh is Vercel's free, open directory for agent skills, launched January 2026, with an MIT-licensed CLI that installs skills via one command: npx skills add owner/repo.
- The same skill works across 20 coding agents, including Claude Code, Cursor, Codex, GitHub Copilot, Windsurf and Zed.
- skills.sh lists 83,627 skills with 8M+ total installs; its Official section lists 6,588 skills from 102 organizations.
- Every indexed skill is scanned by Gen Agent Trust Hub, Socket and Snyk, but skills.sh states it cannot guarantee the quality or security of every listed skill.
- SkillzBench is AIQ Labs' independent testing lab, currently pre-launch, being built to pin every skill version to a commit SHA, security-scan it, and benchmark it in a sandbox.
- SkillzBench will publish a Lift metric — the percentage-point improvement with the skill versus without, using the same model — kept separate from its 0-100 trust score, and money never touches scores.
- Published research found roughly 26% of skills carry at least one vulnerability, which is the gap SkillzBench is being built to close.
Head to head
| Dimension | skills.sh | SkillzBench |
|---|---|---|
| Pricing | Free directory; MIT-licensed CLI; no subscription listed 🏆 | Pre-launch; no public pricing yet |
| Ease of use | One-command npx install, no sign-in needed 🏆 | Planned MCP server plus per-agent loadouts (in development) |
| Discovery | Leaderboard with Hot/Trending/Top, 83,627 skills, 8M+ installs 🏆 | Continuous GitHub crawling and deduplication (in development) |
| Security review | Partner scans (Gen Agent Trust Hub, Socket, Snyk); no guarantee of every listed skill | Commit-SHA pinning, validation, license checks, security scans per version (in development) |
| Performance evidence | Install counts and trending momentum only | Sandbox benchmark with same model, with/without skill, published Lift in percentage points (in development) 🏆 |
| Agent compatibility | 20 agents: Claude Code, Cursor, Codex, Copilot, Windsurf, Zed and more 🏆 | Targets Claude Code, Codex, Cursor, Gemini CLI and chat-app agents via one MCP server |
| Team controls | None: no orgs, roles, access control or audit trail | Per-agent loadouts; Teams tier planned with org loadouts, allowlists and audit 🏆 |
| Availability | Live and free today 🏆 | Pre-launch; waitlist/newsletter live, product not yet public |
The contenders
skills.sh is Vercel's open directory and leaderboard for AI agent skills, launched in January 2026. Skills are modular packages of procedural knowledge — a SKILL.md file plus optional scripts, templates and assets — that teach coding agents how to handle specific tasks without bloating the context window. The MIT-licensed CLI installs any listed skill with one command (npx skills add owner/repo), auto-detects which agent you're using, and drops the skill into the right directory. The same skill works across 20 agents including Claude Code, Cursor, Codex, GitHub Copilot, Windsurf and Zed. The directory ranks skills by install count and trending momentum, with sections for Hot, Trending, Top, Packs, Topics, Official and Audits. As of early 2026 it listed 83,627 skills with 8M+ total installs, and the Official section alone held 6,588 skills from 102 organizations including Vercel, Anthropic, Microsoft and Google.
pricing 0
- + Completely free with no sign-in required to install skills
- + Genuinely cross-agent: one skill installs into 20 different coding agents
- + Simple, inspectable Markdown-based format with an open-source MIT CLI
- + Real adoption signal: 8M+ installs and per-agent breakdowns, with hourly deduplication against replayed installs
- + Partner security scanning (Gen Agent Trust Hub, Socket, Snyk) and an Audits section surface audit information
- + Selective installs, one-time use, and lifecycle commands give fine-grained control
- − skills.sh states it cannot guarantee the quality or security of every listed skill
- − Skills can contain executable scripts and tool permissions, so installing one is a software supply-chain decision, not harmless prompt copying
- − Popularity is based on anonymous install telemetry and does not prove correctness, maintenance quality, or fit
- − Everything is public: no natural home for private internal skills, no version pinning at publish time, no review gate before a skill reaches developers
- − No team constructs: no organizations, member roles, access control, or audit trail
- − Agent support varies — a skill's instructions and scripts may not work identically across every client
SkillzBench is AIQ Labs' independent testing lab and control panel for AI agent skills, currently in pre-launch and actively being built. It addresses the trust gap in the skills ecosystem: published research found roughly 26% of skills carry at least one vulnerability, and catalog size tells you nothing about whether a skill helps or harms. SkillzBench is building five pillars: Discover (continuously crawl and dedupe skills from GitHub), Verify (pin every version to a commit SHA, validate, license-check, security-scan and quality-judge), Benchmark (test every skill in a sandbox with the same model, with and without the skill, and publish the Lift in percentage points), Equip (one MCP server plus per-agent loadouts, where users toggle skills and tools per agent and the agent only sees what was chosen), and Stay current (track upstream changes, re-scan, re-benchmark and notify). Every skill gets a 0-100 trust score built from security, quality, license, maintenance and community signals; Lift is always shown separately from ratings and money never touches scores. Launch scope is instruction-only skills (SKILL.md plus text references); script-bearing skills will be listed link-out only.
pricing See pricing
- + Independent verification: every skill version pinned to a commit SHA, license-checked, security-scanned and quality-judged
- + Measured impact, not popularity: sandbox benchmarks with the same model produce a published Lift in percentage points
- + Per-agent loadouts via one MCP server mean an agent only sees the skills and tools you chose for it
- + Stay-current tracking re-scans and re-benchmarks skills when upstream changes, instead of leaving stale installs behind
- + Trust scores are firewalled from commercial influence — money never touches scores
- + Addresses a documented problem: published research found ~26% of skills carry at least one vulnerability
- − Pre-launch: the waitlist/newsletter is live but the product itself is not yet public
- − Launch scope is instruction-only skills (SKILL.md plus text references); script-bearing skills will be link-out only
- − No live user counts, live skill counts, or published benchmark results exist yet
Verdict
Winner: none
skills.sh and SkillzBench solve different halves of the same problem: skills.sh is a free discovery-and-install layer live today, while SkillzBench is a pre-launch verification and benchmarking layer. They are complements more than substitutes.
| Option | Overall | Breakdown |
|---|---|---|
| skills.sh | 7.5 | : 10: 9: 8: 5: 3 |
| SkillzBench | 7 | : 7: 6: 7: 9: 2 |
Trust and safety compared
Performance evidence: popularity vs measured Lift
Workflow and team fit for small businesses
Which should you pick?
- :
- :
- :
- :
- :
FAQ
Is skills.sh free to use?
Yes. Browsing the skills.sh directory and using Vercel Labs' MIT-licensed skills CLI have no listed subscription fee, and installing skills requires no account — only the public API is rate-limited. Costs can still arise from the AI agent, model usage, or external services an installed skill invokes, which are separate from skills.sh itself.
How many agents does skills.sh support?
skills.sh lists 20 compatible coding agents, including Claude Code, Cursor, Codex, GitHub Copilot, Windsurf, Gemini, Cline, VS Code and Zed. The same skill installs into each of them, though a skill's instructions, scripts and activation behavior may not work identically across every client.
Is SkillzBench available yet?
No. SkillzBench is pre-launch and actively being built; the waitlist and newsletter are live but the product itself is not yet public. Its five pillars — Discover, Verify, Benchmark, Equip and Stay current — describe what SkillzBench is building and will offer at launch, which begins with instruction-only skills (SKILL.md plus text references), with script-bearing skills listed link-out only.
Does skills.sh guarantee the skills it lists are safe?
No. skills.sh states it cannot guarantee the quality or security of every listed skill. It does run partner security scanning — every indexed skill is scanned by Gen Agent Trust Hub, Socket and Snyk, and skills failing every partner audit are excluded — but inclusion is not a security guarantee, and installing a skill with scripts is a supply-chain decision that warrants review.
Can the two tools be used together?
Yes, and that's the most practical path for many small businesses. skills.sh solves discovery and installation today for free, while SkillzBench is being built to add the verification layer the ecosystem lacks — commit-SHA version pinning, security scanning, and measured Lift benchmarks. Discover on skills.sh, then verify and control via SkillzBench once it launches.
Why does skill popularity not equal quality?
skills.sh ranks by anonymous install telemetry, which shows adoption but not correctness, maintenance quality, security, or fit for your project. Ecosystem research reinforces the point: an audit of 22,511 skills found 140,963 issues, Snyk detected prompt injection in 36% of skills tested, and roughly 26% of skills carry at least one vulnerability — popularity-based ranking cannot surface any of that.