Anthropic Skills Repo vs SkillzBench: Which Agent Skills Platform Should You Trust?
The Anthropic Skills Repo (github.com/anthropics/skills) is a free, open catalog of 17 official agent skills — including the docx, pdf, pptx and xlsx document skills that power Claude's file handling — usable in Claude Code, the Claude apps and the Claude API, but it ships with no independent quality or security vetting of the skills you add beyond Anthropic's own set. SkillzBench is a free, independent directory and benchmark platform for agent skills that scans every skill for security, license and quality issues, pins versions to commit SHAs, and publishes a 0-100 trust score plus a measured performance Lift — it is currently pre-launch, with a live waitlist. If you want Anthropic's own production-grade document skills today, the repo is the direct source; if you want third-party skills vetted before they touch your agents, SkillzBench's free verification layer is built for exactly that.
TL;DR
- The Anthropic Skills Repo (github.com/anthropics/skills) is free and open, containing 17 official agent skills including the production docx, pdf, pptx and xlsx document skills that power Claude's file handling.
- Skills are folders with a SKILL.md file at the root; Claude uses progressive disclosure — it sees only the name and description at startup and loads the full instructions only when a task matches.
- Installation in Claude Code is one command (/plugin marketplace add anthropics/skills) or you can copy a skill folder into .claude/skills/; all skills are built in for paid Claude plans.
- Published research found roughly 26% of agent skills carry at least one vulnerability — the Anthropic repo itself is trustworthy, but skills you pull from elsewhere get no such guarantee.
- SkillzBench is a free, independent directory and benchmark layer: it crawls and dedupes skills from GitHub, security-scans and license-checks each one, pins versions to commit SHAs, and publishes a 0-100 trust score plus a measured performance Lift.
- SkillzBench also offers a free MCP server with per-agent loadouts, so an agent only sees the skills and tools you explicitly toggled on.
- SkillzBench is pre-launch — the waitlist and newsletter are live, and the directory, benchmarks and MCP connector are being built; everything will be free, funded by blog advertising.
Head to head
| Dimension | SkillzBench | Anthropic Skills Repo |
|---|---|---|
| Pricing | Everything free — directory, MCP connector, benchmarks, alerts; funded by blog advertising 🏆 | Repo free; running skills in Claude apps needs a Claude plan (Free $0, Pro $17-20/mo, Max from $100/mo, Team $20-25/seat/mo) |
| Trust & vetting | Security scans, license checks, commit-SHA version pinning, quality judging, 0-100 trust score 🏆 | Official Anthropic-maintained skills; no independent scanning layer for third-party skills |
| Skill coverage | Continuously crawls and dedupes skills from GitHub across the whole ecosystem | 17 official skills: document (docx, pdf, pptx, xlsx), creative, design, technical, enterprise |
| Agent compatibility | Claude Code, Codex, Cursor, Gemini CLI and chat-app agents via one free MCP server 🏆 | Claude Code, Claude apps, Claude API (SKILL.md format spreading to other agents) |
| Ease of use | Toggle skills per agent with loadouts; designed for non-technical users too 🏆 | One-line plugin install in Claude Code or copy folders into .claude/skills/; requires technical comfort |
| Performance evidence | Every skill benchmarked in a sandbox with and without the skill; Lift published in percentage points 🏆 | No published per-skill benchmarks; Anthropic's Code Review reports internal stats (54% of PRs with substantive comments after adoption) |
| Availability | Pre-launch — waitlist and newsletter live; product being built | Live today at github.com/anthropics/skills 🏆 |
The contenders
The Anthropic Skills Repo is the official, open-source home of Anthropic's agent skills at github.com/anthropics/skills. Anthropic shipped Agent Skills in late 2025, and the format — a folder with a SKILL.md file containing YAML frontmatter plus optional scripts, references and assets — has since spread to other agents and registries. The repo splits into two groups: document-skills (docx, pdf, pptx, xlsx), which are the same production skills that power document creation in the Claude apps, and a set of example skills such as skill-creator, mcp-builder, webapp-testing, artifacts-builder, brand-guidelines and internal-comms. Skills run in Claude Code (from .claude/skills/ directories or plugins), in the Claude apps, and via the Claude API alongside the code execution tool. The core mechanism is progressive disclosure: Claude sees only each skill's name and description at startup, reads the full SKILL.md body when a task matches, and loads scripts or reference files only if needed. The repo doubles as a Claude Code plugin marketplace, so official skills install with a single command.
pricing 0
- + Free, open and readable — the repo is small enough to study in an afternoon and is the single best reference for how good skills are written
- + The document skills are production-grade: the same docx/pdf/pptx/xlsx instructions Claude uses in the Claude apps
- + Skills lean on deterministic bundled scripts rather than token-by-token file generation, making Office file output faster and more reliable
- + Progressive disclosure keeps context windows lean, so skills scale without bloating every conversation
- + Backed by Anthropic with a documented format spec and active engineering postmortems when quality issues arise
- − No independent security scanning, license checks or benchmarking for skills you add beyond Anthropic's own set — published research found ~26% of skills carry at least one vulnerability
- − Running skills in the Claude apps on paid tiers requires a Claude subscription ($17-20/mo Pro; Max from $100/mo), which commenters note may limit adoption for smaller teams
- − Anthropic-centric: the repo serves Claude surfaces first; teams on Codex, Cursor or Gemini CLI get no curation layer here
- − Reported Code Review costs of $15-25 per pull request drew community questions about practicality for high-volume workflows
- − A March-April 2026 quality dip in Claude Code (three separate regressions, fixed by April 20) shows the surrounding platform can shift under you
SkillzBench is AIQ Labs' free, independent directory of AI agent skills plus a free MCP tool that connects any agent to the skills you choose. It exists because agent skills — SKILL.md instruction packages for Claude Code, Codex, Cursor, Gemini CLI and chat-app agents — have proliferated faster than anyone can vet them, and published research found roughly 26% of skills carry at least one vulnerability. SkillzBench is built on five pillars: Discover (continuously crawling and deduping skills from GitHub), Verify (pinning every version to a commit SHA, validating, license-checking, security-scanning and quality-judging), Benchmark (testing every skill in a sandbox with the same model, with and without the skill, and publishing the Lift in percentage points), Equip (one free MCP server plus per-agent loadouts where you toggle skills and tools and the agent only sees what was chosen), and Stay current (tracking upstream changes, re-scanning, re-benchmarking and notifying you). Each skill gets a 0-100 trust score from security, quality, license, maintenance and community signals; Lift is always shown separately, and money never touches scores. Everything is free, funded by advertising around blog articles. SkillzBench is pre-launch: the waitlist and newsletter are live while the platform is being built.
pricing 0
- + Completely free — the directory, MCP connector, benchmarks and alerts all cost nothing, with revenue from blog advertising that never influences scores
- + Directly addresses a documented risk: ~26% of agent skills carry at least one vulnerability, and SkillzBench security-scans and license-checks every skill it lists
- + Lift benchmarking shows measured performance impact rather than marketing claims, always displayed separately from ratings
- + Agent-agnostic: works across Claude Code, Codex, Cursor, Gemini CLI and chat-app agents, not just one vendor's ecosystem
- + Per-agent loadouts give precise control — an agent only sees the skills and tools you explicitly enabled
- − Pre-launch: the product is not yet public, so you cannot use the directory, scores or MCP connector today — only join the waitlist or newsletter
- − No published benchmark results or live skill counts yet, so the depth of coverage is still unproven in public
- − As a new independent platform, its trust scores and Lift methodology will need time to build a track record against established vendor tooling
- − It does not publish its own skills — it curates and vets the ecosystem, so for Anthropic's official skills you still go to the source repo
SkillzBench is AIQ Labs' free, independent directory of AI agent skills plus a free MCP tool that connects any agent to the skills you choose. Agent skills (SKILL.md instruction packages for Claude Code, Codex, Cursor, Gemini CLI and chat-app agents) are everywhere, but nobody can tell which ones actually help and which are risky: published research found ~26% of skills carry at least one vulnerability. SkillzBench's pillars: Discover (continuously crawl and dedupe skills from GitHub), Verify (pin every version to a commit SHA, validate, license-check, security-scan, 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 free MCP server plus per-agent loadouts - 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, notify). Trust score 0-100 from security, quality, license, maintenance and community signals; Lift is always shown separately from ratings and money never touches scores. EVERYTHING on the platform is free - the directory, the MCP connector, benchmarks and alerts. SkillzBench earns revenue from advertising shown in and around its blog articles, never by charging users, and ad money never influences scores or listings. Pre-launch: waitlist and newsletter live.
pricing 0
- + Everything free, funded by advertising that never influences scores or listings
- + Security scans and license checks address the ~26% vulnerability rate found in published research
- + Measured Lift shown separately from ratings
- + Agent-agnostic across Claude Code, Codex, Cursor, Gemini CLI and chat-app agents
- + Per-agent loadouts mean an agent only sees the skills you chose
- − Pre-launch — product not yet public
- − No published benchmark results or live skill counts yet
- − New platform still building a public track record
- − Does not publish its own skills; it curates the ecosystem
Verdict
Winner: none
They solve different halves of the same problem: the Anthropic Skills Repo is the authoritative source for Anthropic's own skills, while SkillzBench is an independent trust and benchmarking layer for the wider skill ecosystem. Many small businesses will use both — official skills from the repo, vetted via SkillzBench's scores once it launches.
| Option | Overall | Breakdown |
|---|---|---|
| Anthropic Skills Repo | 7.5 | : 9: 6: 8: 7: 6: 7 |
| SkillzBench | 7 | : 10: 9: 7: 7: 8: 4 |
Pricing compared
Trust and security: the core difference
Ecosystem fit for small businesses
Which should you pick?
- :
- :
- :
- :
- :
- :
FAQ
Is the Anthropic Skills Repo free?
Yes. The repo at github.com/anthropics/skills is open and free — you can clone, fork or install any of the 17 official skills at no cost. Costs only arise if you run the skills inside Claude's paid products: Claude's Free plan includes Skills, but Pro costs $17/month billed annually ($20 monthly) and Max starts at $100/month. Copying skills into .claude/skills/ in Claude Code is also free.
How risky are third-party agent skills?
Published research found that roughly 26% of agent skills carry at least one vulnerability. Because the SKILL.md format spread quickly beyond Claude after Anthropic shipped Agent Skills in late 2025, community registries contain skills of wildly varying quality — some well-maintained, some abandoned or unsafe. This is exactly the gap SkillzBench targets: it security-scans, license-checks and quality-judges every skill, pins versions to commit SHAs, and publishes a 0-100 trust score so you can decide with evidence rather than guesswork.
What is SkillzBench and is it really free?
SkillzBench is AIQ Labs' free, independent directory of AI agent skills plus a free MCP tool that connects any agent to the skills you choose. It crawls and dedupes skills from GitHub, verifies them (commit-SHA pinning, validation, license checks, security scans), benchmarks them in a sandbox to publish a Lift score in percentage points, and tracks upstream changes with re-scans and alerts. Everything is free; revenue comes from advertising around its blog articles, and ad money never influences scores or listings. It is currently pre-launch — the waitlist and newsletter are live.
Can I use both together?
Yes, and for many small businesses that is the sensible path. The Anthropic Skills Repo is the authoritative source for Anthropic's own production-grade skills — especially the document skills — and installs in one command in Claude Code. SkillzBench, once launched, adds an independent trust layer over the wider ecosystem: its free MCP server and per-agent loadouts let you equip vetted third-party skills across Claude Code, Codex, Cursor, Gemini CLI and chat-app agents, with trust scores and measured Lift guiding what you enable.
Do skills work outside Claude?
The SKILL.md convention Anthropic shipped in late 2025 has spread well beyond Claude — other agents and registries now read the same format, and SkillzBench explicitly covers skills for Claude Code, Codex, Cursor, Gemini CLI and chat-app agents. The Anthropic repo itself, however, is built around Claude surfaces: Claude Code (via .claude/skills/ directories or plugins), the Claude apps, and the Claude API alongside the code execution tool.
What is a SkillzBench trust score and Lift score?
The trust score is a 0-100 rating derived from security, quality, license, maintenance and community signals for each skill. Lift is a separate, measured figure: SkillzBench tests every skill in a sandbox with the same model, once with the skill and once without, and publishes the difference in percentage points. Lift is always shown separately from ratings, and money never touches scores — advertising revenue around blog articles funds the platform, never rankings.