versuz

Versuz

Skills go in. Only one wins.

The open public benchmark for AI agent skills.

Versuz indexes every public Claude skill and CLAUDE.md on GitHub, runs them through 5 held-out tasks, lets three frontier LLMs judge the outputs, and publishes a ranking based on actual performance, not stars.

Free · Open · Updated daily

🌐 versuz.dev · 📦 npx versuz · 🔌 @versuz/mcp · 📚 Methodology

Why it exists

In May 2026, the Claude Code ecosystem already lists 4,200+ skills on claudemarketplaces.com and 1.2M+ on skillsmp.com. Anthropic has shipped its own enterprise marketplace. None of them answer the only question a developer actually asks :

"Of the 47 skills claiming to do task X, which one actually works?"

Existing directories rank by stars and installs — popularity, not quality. Versuz benchmarks the outputs. Three frontier judges grade every skill against the same task suite, and the winners rank #1 in their category.

It's LMArena, but for agent skills.

What's indexed

Three commercial tiers

• Free — scraped public, verified progressively (5-level trust ladder : claimed → verified → reviewed → featured)
• Premium — author-listed, fixed price, Versuz takes 30% / author keeps 70% via Stripe Connect Express
• Featured — Versuz first-party curation, 100% Versuz

Consume Versuz from anywhere

Web UI

versuz.dev/marketplace — full registry with filters (kind, category, tier, trust level, quality, tokens, official, topic).

CLI (v0.2 — beta)

npx versuz                              # interactive mode
npx versuz search pdf                   # cross-kind full-text search
npx versuz install <slug>               # download to .claude/skills/<slug>/
npx versuz battle <a> vs <b>            # head-to-head terminal viz (v0.2)
npx versuz login                        # auth with GitHub PAT
npx versuz submit <github-url>          # share your skill (own repo only)

→ Full CLI docs

MCP server (v0.1 — beta)

claude mcp add versuz npx -y @versuz/mcp

Claude Code now has 5 native tools : versuz_search, versuz_list_skills, versuz_list_claude_md, versuz_get, versuz_install. Ask "find me a PDF skill and install it" — Claude searches, inspects, downloads, all inline.

→ Full MCP docs

Open data

• GET /api/v1/skills · GET /api/v1/skills/<slug> · /content
• GET /api/v1/claude-md · same shape
• GET /feed/skills · GET /feed/claude-md — RSS 2.0
• GET /sitemap.xml
• Per-item embed badges : /badge/<kind>/<slug> (SVG)

How it works (short)

1. Discover — multi-source scrape (GitHub Code Search + Sourcegraph + awesome-* aggregators) with multi-token rotation, content-hash dedup, official-org auto-flag.
2. Score — LLM 5-axis quality judge (clarity, specificity, completeness, structure, usefulness) on every item. Mean 67/100, target distribution N(65, 12).
3. Benchmark — for each category, every skill runs through 5 held-out tasks (N=5 = statistically validated sweet spot). Output dedup by content hash so identical runs are reused.
4. Judge — 3 frontier LLMs grade each output independently. Judge dedup by output hash. Prompt cache hits ~82%, cutting input cost by ~57%.
5. Aggregate — Bayesian Elo per category. Refreshed every 15 min via Vercel cron.

Full methodology : versuz.dev/methodology.

Stack

Local dev

git clone https://github.com/TomaTV/versuz
cd versuz
npm install
cp .env.local.example .env.local   # fill Supabase, Stripe, Resend, GitHub PATs
npm run dev                        # http://localhost:3000

Useful scripts

Env vars (essentials)

NEXT_PUBLIC_SUPABASE_URL=
NEXT_PUBLIC_SUPABASE_ANON_KEY=
SUPABASE_SERVICE_ROLE_KEY=
GITHUB_TOKENS=ghp_xxx,ghp_yyy,ghp_zzz   # multi-token rotation
OPENROUTER_API_KEY=                     # bench mode or-v1
STRIPE_SECRET_KEY=
STRIPE_WEBHOOK_SECRET=
RESEND_API_KEY=
ADMIN_GITHUB_LOGINS=YourGitHubLogin

Repo layout

versuz/
├── src/                     # Next.js app + lib + components
│   ├── app/                 # App Router pages + /api routes
│   ├── components/          # Brand, site, marketplace, motion, …
│   └── lib/                 # queries, auth, stripe, supabase, …
├── scripts/                 # Node scrapers + bench engine
│   ├── scrape/              # GitHub Code Search (SKILL.md)
│   ├── scrape-claude-md/    # GitHub Code Search (CLAUDE.md)
│   ├── scrape-aggregators/  # 14 awesome-* + 8 topics
│   ├── scrape-codesearch/   # Sourcegraph adapter
│   └── bench/               # Agent + judge + cycle orchestrator
├── supabase/migrations/     # 22 SQL migrations
├── cli/                     # `npx versuz` CLI package
├── mcp-server/              # `@versuz/mcp` MCP server package
├── docs/                    # Go-live checklists (stripe, domain, cli-mcp)
├── .ui/                     # Design reference (Claude Design v1)
├── CONTEXT.md               # Long-form project context
├── CLAUDE.md                # Agent working agreement
└── TODO.md                  # Task list

Bench cost calculator

Mode or-v1 (default) : Haiku 4.5 + DeepSeek V3 + GPT-5 mini via OpenRouter. One key, one dashboard.

Why DeepSeek V3 over V4 Flash : V4 Flash is reasoning-only and we disable reasoning to save tokens, which polarized its scoring (stddev 27, 46% sub-30). V3 chat is non-reasoning by design — stable calibration (avg 57, stddev 14) for a small cost bump (~$0.20/day at 100 skills).

Optimisations available :

• BENCH_JUDGE_COUNT=2 → cut judge calls by 33%, save ~$27 on 5k skills
• BENCH_JUDGE_MAX_TOKENS=500 → squeeze DeepSeek reasoning further, save ~$15 on 5k skills
• BENCH_MODE=or-thrift → single judge (GPT-5 nano), ~$0.86/day
• BENCH_MODE=dev → 3 free Groq Llama models, $0 (3000 RPD cap)

Roadmap

• ✅ V0 — public marketplace, scraping pipeline, quality judge, CLI v0.1, MCP v0.1
• ✅ V0.5 — bench engine live (or-v1 with prompt cache), first cycles completed
• ✅ V1 — domain launch, Stripe live mode, full catalog bench
• ✅ V1.6 — Cloudflare R2 migration, gamification (achievements + streaks), Today's Upset content automation, 5 monetization surfaces, CLI v0.2 (battle command), bench-tier filter
• 🔮 V2 — real-time battles (LMArena-style live judging), dark theme, API plans for companies

→ Full plan in TODO.md.

Built by

FlukX Studio — Toma, solo build from France. EdTech automation + design + full-stack.

License

MIT — see LICENSE.