The bottom line

Start with Anthropic's own repository, then use a registry for breadth and a curated list for judgment. Nobody has solved discovery for Agent Skills yet. The catalogs that are big are automatically scraped and cannot tell you what is good; the lists that exercise taste are hand-maintained and cannot cover the field. Using one of each is not a workaround, it is currently the correct approach.

Every option below is free. None of them have paid tiers, and ToolChase has no commercial relationship with any of them.

What a skill actually is

A skill is a markdown file with YAML frontmatter that teaches an agent a workflow or a domain: instructions, conventions, examples. It contains no code and runs no process. You place it in your agent's skills directory and it changes how the agent behaves.

The most common confusion worth clearing up first: skills are not MCP servers. A skill changes how the agent works. An MCP server lets the agent reach a live system, such as a database or a ticket tracker. Teaching the agent your team's commit conventions is a skill. Letting it read your actual Jira board needs an MCP server. They live in separate registries, and searching the wrong one is the usual reason people decide nothing exists for their use case.

Because the format is plain markdown, skills are portable across Claude Code, Codex, GitHub Copilot and other assistants, and you can always fork one into your own repo and stop depending on any registry.

The official skills, and what each one is actually for

Anthropic publishes 19 skills in its own repository, and most write-ups mention four of them. The document skills are the famous ones because they power Claude's file creation, but they are also the only four in the set that are source-available rather than open source. The other 15 are Apache 2.0, which matters if you intend to fork one and ship it inside your own product.

Anthropic's own framing is that these are demonstrations, not a supported product surface. Treat them as reference implementations you read and adapt, and test anything you plan to rely on.

Document skills: docx, pdf, pptx, xlsx

These four cover Word documents, PDFs, presentations and spreadsheets, and they are the ones actually running behind Claude's document features. Scope is wider than "make a file": the PDF skill also merges, splits, rotates, extracts text and tables, fills forms and runs OCR over scans, and the spreadsheet skill will repair a messy CSV as readily as build a new workbook. If you only install one group, install this one. Note the license difference above before reusing the code.

Development and technical: mcp-builder, skill-creator, webapp-testing, claude-api, doc-coauthoring

mcp-builder is the one to reach for when a skill is the wrong answer, because it walks through building an MCP server in Python or TypeScript to connect a real external service. skill-creator is the meta-skill: it writes new skills, improves existing ones, and runs evals to measure whether a change actually helped, which is the part most hand-rolled skills skip. webapp-testing drives a local web app through Playwright to verify behavior, capture screenshots and read browser logs. claude-api is a working reference for the Claude API and SDK covering model ids, pricing, streaming, tool use, caching and migration. doc-coauthoring structures the back and forth of writing a spec or a decision doc rather than producing one in a single shot.

Design and creative: frontend-design, canvas-design, theme-factory, algorithmic-art, web-artifacts-builder, slack-gif-creator

This group is the most underused, and it is the reason to look past the document four. frontend-design pushes for deliberate typography and aesthetic direction instead of the templated defaults that make generated UI recognizable on sight. canvas-design produces static pieces such as posters as PNG or PDF. theme-factory applies one of ten preset themes, or a generated one, across slides, docs and landing pages so a set of artifacts looks like a set. algorithmic-art generates work in p5.js with seeded randomness, so a result you like is reproducible. web-artifacts-builder targets multi-component HTML artifacts using React, Tailwind and shadcn/ui, for cases where a single file is not enough. slack-gif-creator is narrow and unusually practical: animated GIFs built to Slack's constraints, with validation.

Enterprise and communication: brand-guidelines, internal-comms, academy-guide, discernment-nudge

brand-guidelines applies Anthropic's own colors and typography to an artifact, and is most useful read as a worked example of encoding a brand system, since you will want your own. internal-comms covers status reports, leadership updates and incident write-ups in house formats. academy-guide points questions about using Claude toward matching courses on Claude Academy. discernment-nudge is the most interesting of the four and the least like a normal skill: after a substantive answer it appends two or three follow-up questions aimed at checking facts and surfacing assumptions, at most once per conversation. It is a guard against acting on a confident draft, which is a real failure mode rather than a feature.

How to install them in Claude Code

The repository doubles as a Claude Code plugin marketplace, so you do not have to copy folders by hand. Register it once:

/plugin marketplace add anthropics/skills

Then install either bundle directly:

/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

After installing, invoke a skill by mentioning it, for example asking Claude Code to use the PDF skill to extract form fields from a specific file. On Claude.ai these example skills are already available to paid plans, so there is nothing to install there. The repository also ships the Agent Skills specification and a blank template, which are the two things to read before writing your own.

1. Anthropic's official skills repository, the canonical reference

github.com/anthropics/skills is the first-party repository, and it is the right place to begin even though it is by far the smallest option here. It holds a deliberately limited set of demonstration skills across creative, technical and enterprise work, plus the format specification in /spec and a starter template in /template. The document skills (docx, pdf, pptx, xlsx) are the same family that powers Claude's own document handling.

In Claude Code it installs with one command: /plugin marketplace add anthropics/skills. Most skills carry Apache 2.0; the document skills are source-available rather than open source, which Anthropic states plainly.

Use it for: learning what a well-formed skill looks like before you try to evaluate a stranger's. Limitation: it is a reference set, not a catalog, so you will outgrow it quickly.

2. AgentSkills.codes, the broadest scraped index

AgentSkills.codes indexes 19,296 installable skills across nine categories, Coding being the largest at 6,343. Everything is scraped from public GitHub repositories rather than vendor-submitted, and the site re-scans the whole catalog daily for freshness and safety signals.

The part that earns its place is per-listing provenance: source repo, stars, license and last commit date on the card, so an abandoned skill is obvious before you open it. It is MIT licensed, free, has no accounts and no paid tier, and lists compatibility beyond Claude Code including Codex, GitHub Copilot and Antigravity.

Use it for: answering "has anyone already built this" fast. Limitation: scraping means inclusion proves existence, not quality, and at nearly 20,000 entries the hard part becomes judgment. Our full AgentSkills.codes review sets out the security posture worth adopting before you install anything from it.

3. SkillsMP, aggregation at extreme scale

SkillsMP describes itself as the largest community-driven marketplace for agent skills, and says it indexes more than 2 million open-source skills across 12 domains, 50-plus categories and nine languages. That figure is the vendor's own claim and is an order of magnitude beyond anything else here, which tells you the aggregation is extremely wide rather than filtered.

It is free with no paid tiers, and offers an API: 50 requests a day anonymously, 500 with a free key. Like AgentSkills.codes it pulls from GitHub repositories and exposes search, categories and occupation filters.

Use it for: programmatic access, and for long-tail searches where a smaller index returns nothing. Limitation: at that scale the signal-to-noise problem is at its most acute, and the site does not identify who operates it.

4. awesome-claude-skills, the curated shortlist

ComposioHQ/awesome-claude-skills is a hand-picked list rather than a catalog, and that is exactly its value. It carries roughly 72,000 GitHub stars, which is a meaningful signal that a lot of practitioners have found it useful. Its most recent update shipped in late July 2026.

Use it for: the judgment layer the scraped registries cannot provide. Limitation: curation moves at human speed, so recent publications lag, and "awesome list" inclusion still reflects one maintainer's taste. It is maintained by Composio, a commercial company in the agent tooling space, which is worth knowing even though the list itself is open.

5. awesome-claude-code, skills in the context of everything else

hesreallyhim/awesome-claude-code is broader than skills alone: it collects commands, hooks, workflows and configuration patterns for Claude Code generally, with skills as one section. Around 52,000 stars, and it was updated the same week we published this guide, which makes it one of the more actively maintained resources in the space.

Use it for: the case where a skill is not actually what you need and a hook or slash command solves the problem better. Limitation: skills are a subsection, so the depth per category is shallower than a dedicated index.

6. Claude Code Templates, skills bundled with agents and commands

aitmpl.com packages more than 1,000 components spanning skills, agents, commands, settings, hooks and MCP integrations, with a stack builder for assembling a configuration rather than picking files one at a time. It is free and open source, supported by sponsorships from developer-tool companies.

Use it for: setting up a new project configuration end to end instead of hunting individual skills. Limitation: the component counts are self-reported and the browsing experience is dynamic, so it is harder to audit what is actually there than a plain repository listing.

7. claude-code-plugins-plus-skills, plugin-first packaging

jeremylongshore/claude-code-plugins-plus-skills advertises 471 plugins, 3,069 skills and 347 agents, MIT licensed, and was pushed to the same week as publication. It leans on Claude Code's plugin marketplace mechanism rather than manual file placement.

Use it for: installing groups of related skills through the plugin system. Limitation: single-maintainer bundles concentrate risk, so the security reading below matters more here, not less.

How to judge a skill before you install it

This is the part every directory leaves to you, so it is worth being systematic. A skill is instructions your agent will follow, which is a materially different risk profile from a library you import and never read.

• Read the markdown. Skills are short by design. This is the rare case where reading the whole dependency takes two minutes.
• Watch for shell, credentials and deletion. A skill that instructs the agent to run commands, touch secrets or remove files deserves real scrutiny.
• Check the commit date, not just the stars. Star counts are lagging indicators; an unmaintained skill written against an older agent version can quietly misfire.
• Find the upstream original. Scraped catalogs surface forks, so the first result is not always the source. Compare stars and dates.
• Prefer forking to depending. Copy what you use into your own repo. The format is portable precisely so you can.

Which one should you use

If you are new to skills: Anthropic's repository, then a curated list. If you know what you want and need to find it: a scraped registry, AgentSkills.codes for readable provenance or SkillsMP when you need the widest possible net. If you are configuring a project from scratch: Claude Code Templates. In practice most people end up with a curated list bookmarked and a registry open in another tab, because no single source currently does both jobs.

Related reading: our Claude Code review, the best AI coding agents, and Claude Code vs Cursor.

FAQ

What is a Claude Code skill?

A skill is a markdown file with YAML frontmatter that teaches an agent a workflow or a body of domain knowledge. It contains instructions, conventions and examples, not code. You drop it into the agent's skills directory and it shapes how the agent behaves. Because it is just markdown, the same file works across Claude Code, Codex, GitHub Copilot and other assistants that adopted the format.

Are skills the same as MCP servers?

No, and confusing the two sends people to the wrong registry. A skill is instructions that change how the agent works and runs nothing. An MCP server is software you connect so the agent can reach a live system such as a database or a ticketing tool. Team commit conventions are a skill; reading your actual Jira tickets needs an MCP server.

Where should I start if I have never installed a skill?

Anthropic's own repository. It is small, first-party, and shows the format the way it was intended, and it installs in Claude Code with a single plugin marketplace command. Once you know what a good skill looks like, the larger community registries become far easier to judge.

Is it safe to install skills from a public registry?

Apply package-registry caution. A skill is instructions your agent will follow, so a malicious or careless one can direct it to exfiltrate credentials, delete files or run shell commands. The mitigation is unusually easy here: skills are short markdown files, so read one before installing it, and treat scraped catalog inclusion as evidence a skill exists rather than evidence it is any good.

Do these directories cost anything?

Every option in this guide is free at the time of writing. Most are open-source repositories or free registries with no paid tiers and no accounts required to browse. That is normal for an ecosystem this young, and it means the real cost of a bad choice is your time, not a subscription.

Skill counts, star counts, licenses and update dates verified August 2026 from each project's own site or the GitHub API. Star counts move; treat them as a snapshot. ToolChase has no commercial relationship with any resource listed here and none of these links are affiliate links.