knack
← all posts

The Claude Skills Marketplace: Where to Find, Publish, and Share Skills

There is no single Claude skills marketplace. There are several, plus a long tail of skills living in personal GitHub repos that no marketplace has indexed yet. Where to publish, where to shop, and what to check before installing.

There is no single Claude skills marketplace. There are several, plus a lot of skills living in personal GitHub repos that no marketplace has indexed yet.

That fragmentation matters if you are deciding where to publish, or which directory to trust when shopping for something to install. The official anthropics/skills repository on GitHub is the canonical reference. It has seventeen skills in it. Community aggregators like SkillsMP claim more than a million indexed skills, mostly scraped from public GitHub, while claudemarketplaces.com lists around 4,200 skills and 2,500 sub-marketplaces. The numbers do not agree because nobody owns the definition.

This piece walks the current state of the Claude skills marketplace in May 2026, then takes a position on where to publish and what to check before installing.

What a skill actually is, before we shop for one

A Claude skill is a folder. Inside the folder sits a file called SKILL.md with YAML frontmatter at the top (name, description, optional fields like allowed-tools and disable-model-invocation) followed by Markdown instructions. The folder can also contain supporting files: reference docs Claude loads on demand, scripts Claude can execute, templates for it to fill in. That is the entire format, documented at code.claude.com and specified in the open Agent Skills standard under spec/ in the official repo.

Skills live in one of four places on your machine: ~/.claude/skills/ for personal skills available everywhere, .claude/skills/ inside a project repo, inside a plugin, or pushed down by enterprise managed settings. Claude loads the frontmatter into context at session start so it knows what is available, then pulls the body in only when the skill is invoked, either by you typing /skill-name or by Claude deciding the description matches your request. Anthropic calls this progressive disclosure.

That changes what a "marketplace" is for. A skill is a small text artifact you copy into a folder. No runtime, no installer, no sandbox, no payment rail. Distribution is git clone with extra steps.

The official repo, and why it is small on purpose

github.com/anthropics/skills is the closest thing to a first-party Claude skills marketplace. As of this writing it contains seventeen skills, all maintained by Anthropic, grouped loosely into document handling, creative work, development tooling, and internal communications.

A few real ones you can install today:

  • pdf: read, merge, split, rotate, watermark, encrypt, fill forms, run OCR. The frontmatter triggers on any mention of a .pdf file. Body is Python with pypdf, plus reference docs in supporting files.
  • mcp-builder: scaffolds a new MCP server from a description.
  • webapp-testing: drives a headless browser to test a web app end to end.
  • skill-creator: a skill that helps you write other skills, which is funnier than it sounds.
  • slack-gif-creator: builds animated GIFs sized for Slack reactions.

Install one by running /plugin install document-skills@anthropic-agent-skills inside Claude Code, or by cloning the folder into ~/.claude/skills/ directly. Both work. Anthropic recommends the plugin command because it pins to a published version and updates cleanly.

The repo is small on purpose. Anthropic is using it as a reference implementation, not as the place every skill in the world should live. The README points outward to community catalogs for the long tail.

Where the long tail actually lives

Three places, with different shapes.

SkillsMP is an aggregator. It scrapes public GitHub for repositories that look like skills, applies a quality filter (minimum two stars), and presents them through a search UI with occupation and use-case tags. It claims roughly 1.3 million indexed skills, which says more about how many GitHub repos contain a SKILL.md somewhere than about how many useful skills exist. The signal-to-noise is what you would expect when a crawler indexes everything matching a pattern. You will find five different "git commit message generator" skills and read the frontmatter on each to pick one.

claudemarketplaces.com is a directory of directories. It catalogs 4,200 individual skills, 770 MCP servers, and 2,500 sub-marketplaces, where a sub-marketplace is usually a single GitHub repo with a .claude-plugin/marketplace.json file pointing at a set of skills and plugins the owner has curated. Entries worth opening: obra/superpowers (an opinionated agentic framework), f/prompts.chat (community prompts), and the official anthropics/skills itself. This is the most useful starting point if you want to browse, because the curation is human and listings rank by install count and stars rather than by what a crawler found.

Then there is the raw GitHub layer. Many of the most-used skills in May 2026 are not in any catalog. They live in personal repos with a README that says "drop this in ~/.claude/skills/" and a couple hundred stars from word-of-mouth on X and Discord. Anthropic merged custom commands into skills earlier this year, so a lot of the old .claude/commands/*.md files shared on gists are now technically skills, which inflates every count you read.

The fragmentation is the story. A creator in 2026 publishes to GitHub first, then optionally lists in one or two of the aggregators. A consumer searches in two or three places before installing anything they cannot find a thumbs-up for from someone they trust.

How publishing mechanically works

You publish a skill by putting the folder in a public Git repository and pointing people at it. No upload form, no submission queue, no review process at any of the marketplaces I have used. Aggregators index you automatically if the repo is public.

The shape of a publishable skill:

my-skill/
  SKILL.md          # required, with frontmatter
  reference.md      # optional supporting file
  scripts/
    helper.py       # optional script Claude can run

The frontmatter decides whether anyone installs it. Two fields carry real weight. Claude reads description to decide whether your skill matches a request; humans read it on the marketplace listing to decide whether to install. Anthropic caps the combined description and when_to_use text at 1,536 characters before truncation, so put the trigger phrases first. The other load-bearing field is allowed-tools: it grants Claude permission to use the listed tools without per-call approval. A skill that asks for broad tool access (Bash, Write, network) needs to earn the trust, because the consumer is signing off on every one of those at install time.

If you want your skill to show up in a marketplace UI rather than just a search index, you wrap it in a .claude-plugin/marketplace.json file at the root of the repo. That file lists the skills and tells Claude Code how to install them as a plugin. The official anthropics/skills repo uses exactly this pattern, which is why /plugin install document-skills@anthropic-agent-skills works at all.

The annoying part of publishing today is that the domain experts who would write the best skills (lawyers, doctors, accountants, ops people who have run the procedure a hundred times) do not want to learn YAML frontmatter, plugin manifests, and the difference between disable-model-invocation and user-invocable. Tools like Knack fill that gap: a non-coder runs through a guided interview and the tool emits a real SKILL.md folder in the open standard, ready to push to GitHub or share via knack pull <slug>.

If you can write Markdown and edit a YAML block, you do not need any tool. The format is small and the docs are good.

How discovery and installation work on the consumer side

Three install paths, depending on where you found the skill.

From a marketplace UI: click install, and the marketplace either runs /plugin install <name>@<marketplace> for you or copies a one-liner to paste into your terminal. SkillsMP and claudemarketplaces.com both do this. The skill lands in ~/.claude/skills/ and Claude Code picks it up on the next prompt without a restart, because Claude Code watches the skills directory for changes.

From a GitHub URL someone DM'd you: clone the repo, copy the skill folder into ~/.claude/skills/, or use the plugin install command if the repo has a marketplace.json. The plugin path is cleaner because uninstall is a single command and updates pull cleanly.

From a project repo you cloned: skills in .claude/skills/ load automatically once you trust the workspace. Claude Code prompts on first run because a skill with allowed-tools is granting itself permission to use tools without asking, and you should approve that consciously.

Before installing, open the SKILL.md and check three things. First, the description: is it specific, or does it claim to do five unrelated things? Vague descriptions are how skills get auto-triggered when you didn't want them. Second, the allowed-tools line: a skill that pre-approves Bash for arbitrary commands is asking for trust you should not give to a stranger's repo. Third, scan the body and any supporting scripts. If the skill runs a curl | sh or hits a third-party API, that is the moment to decide whether you trust the author. Most skills are short. A five-minute read is enough.

Pricing, and the absence of it

Almost nobody monetizes a skill directly in May 2026. The format has no payment hook, and the marketplaces do not gate downloads behind a paywall. What you see is the pattern that ran the early VS Code extension market: free skills, with the value capture one layer up at the hosted service the skill calls, the consulting engagement the skill leads to, or the SaaS the skill installs and configures. A skill that wires Claude to a paid API still works; the payment lives at the API, not the skill.

The trajectory is obvious. The moment a marketplace bolts on a payment layer and a license format, paid skills will appear, and the first wave will probably look like the paid GPT store circa late 2023. Whether any of them earn real revenue depends entirely on how good the discovery surface is, which today it is not.

A working position for May 2026

If you are publishing, push to GitHub with a marketplace.json so consumers can install via the plugin command, then submit the repo to claudemarketplaces.com for human curation and let SkillsMP pick it up by crawl. Do not wait for any marketplace to grant you access. There is no gate.

If you are installing, browse claudemarketplaces.com first, fall back to SkillsMP search for the long tail, read the SKILL.md before you install, and treat any skill with broad allowed-tools the same way you treat a random shell script someone emailed you.

The marketplace question will probably consolidate inside the next twelve months, once one of the aggregators ships better discovery or Anthropic publishes a first-party directory beyond the reference repo. Until then, the answer is more than one place, and you check each.