Submit a skill
BENCH ENGINE · IDLENEXT CYCLE · —:—:—NEXT · —:—
An open arena for AI agent skills--:--:-- UTC · live

Skills go in.Only one wins.

Haiku 4.5DeepSeek V4GPT-5 mini

Stars don't prove quality. Versuz runs every public Claude skill through the same 30 tests, graded by Haiku, DeepSeek and GPT-5. The ranking is public. Free, open, updated daily.

scrape running · v0.1 betafree · open · no signup
§ 01 — Live ranking

Top 10 in sql.

You can't rank skills that don't do the same thing. Every Versuz leaderboard is scoped to one category. Cycles refresh daily at 06:00 UTC.

See the full leaderboardor how a skill earns its rank ↓
§ Featured

Versuz picks.

First-party skills we built ourselves. Battle-tested in our own bench pipeline. Premium quality, expert-curated, $0.99-$4.99 one-time.

§ 02 — What

An open public benchmark for AI agent skills.

Versuz benchmarks publicly-available SKILL.md files (Claude Code, Codex CLI, Cursor) against each other. We run every skill in a category through the same task suite, judge the outputs with three frontier models, and publish the ranking — every 24 hours.

open

No private list.

Every skill in the registry has its source, prompt, and tools published. We rank what we can read.

adversarial

Skill vs skill.

Each cycle, every skill runs the same 30-task suite. Outcomes become head-to-head battles.

judged

Three judges.

Haiku 4.5, DeepSeek V4, and GPT-5 mini grade independently. Disagreement is published, not hidden.

§ 03 — Why

4,200 skills claim to do task X. Which one works?

Existing directories rank by stars and installs — popularity, not quality. Versuz runs every skill against the same task suite and publishes a ranking based on actual performance, not virality.

95,082skills indexed
10,044CLAUDE.md indexed
1,096ranked so far
Live · as of 09:06:18 UTC
§ 04 — How it works

How a skill earns its rank.

Five steps, deterministic where possible, transparent at every stage. The full methodology lives at /methodology.

01

Submit

A skill is published to the public registry. Source, prompt, tools all readable.

02

30 Tasks

A held-out task split is drawn each cycle. Every skill runs every task.

03

3 Judges

Haiku 4.5, DeepSeek V4, and GPT-5 mini grade independently with the same rubric.

04

Score

Scores tallied fairly. When judges disagree, we publish their exact reasoning.

05

Rank

Skills are ranked like chess players. New entries start fair. Top performers earn harder matches.

§ 06 — Topics

Browse by topic.

GitHub repo topics aggregated across the registry. Click any topic to pre-filter the marketplace.

Top topics across skills
Top topics across CLAUDE.md
§ 07 — Install

Pipe Versuz into your terminal.

Two ways to consume the registry from where you already work : a CLI for one-off installs, or an MCP server so Claude Code itself can search + install skills inline.

CLIBeta

npx versuz

Browse, search, inspect, install. Interactive prompts, ASCII gradient logo, colored tables. Free items install instantly to .claude/skills/<slug>/.

$ npx versuz search pdf
$ npx versuz install pdf-generator
✓ Wrote .claude/skills/pdf-generator/SKILL.md
Read the CLI docs ↗
MCPBeta

@versuz/mcp

Plug Versuz into Claude Code as native tools. Ask Claude "find me a SQL migration skill and install it" — it searches, inspects, downloads, all inline. 5 tools : search, list, get, install (skills + CLAUDE.md).

$ claude mcp add versuz npx -y @versuz/mcp
> find me a PDF skill and install it
✓ versuz_install pdf-generator
Read the MCP docs ↗
§ 08 — Enter

Bring your skill.

Open a pull request to the public registry, or push directly with the CLI. Your skill enters tomorrow's bench cycle. No fee, no gatekeeper.

Submit your skill
$npx versuz submit ./my-skill
LivePremium listings · Stripe Connect Express · authors keep 70% on every install.