Plugins

A plugin is a Claude Code package that bundles a project's skills, agents, hooks and MCP servers so someone else can install the whole setup in one step.

Last updated: 29 Sep, 2026 · Claude Code

By now this project has a CLAUDE.md, a command, a skill, a hook and a subagent. A plugin is that pile in one directory, installable by somebody else.

A plugin is a directory holding skills, agents, hooks or MCP servers, with a manifest that gives it a name.

{
  "name": "link-shortener-kit",
  "description": "Conventions, review skill and protected-file hook for this project",
  "version": "0.1.0"
}

Around that manifest sit the same files you have already written, in the same shapes: skills/, agents/, hooks/. Nothing is rewritten to become a plugin. That is why the documentation suggests starting in .claude/ and converting once it works.

Why bother

• A team gets the same setup. One install instead of a folder of files copied by hand.
• It is versioned. An improvement to the review skill reaches everybody.
• It travels. The same kit works in every repository that installs it.

claude plugin list

No plugins installed. Use `claude plugin install` to install a plugin.

That is a machine with nothing installed. With plugins present each one is listed with its version, its scope, and whether it is enabled.

Plugins are distributed through marketplaces, which are repositories that list plugins. Your team's marketplace can be a private repository, and installing from it is the same command as installing anything else.

Installing Context7 from the official marketplaceOptional

The video opens /plugin. Its Discover tab lists plugins from Anthropic's official marketplace, claude-plugins-official, with their install counts: superpowers, context7 ("Upstash Context7 MCP server for up-to-date documentation lookup"), code-review, code-simplifier and skill-creator among them. It installs context7, and Claude Code answers "Installed context7. Run /reload-plugins to apply."

After the reload the video asks "based on the recent langchain documentation explain how to create deep agents using langchain". The first attempt goes to web search; adding "using context7 plugin" to the prompt makes Claude load the Context7 tools, resolve the LangChain library and query its current docs, and the answer on deep agents and their planning tools comes from them.

To install it yourself, from your shell:

claude plugin install context7@claude-plugins-official

Or run /plugin install context7@claude-plugins-official inside a session and pick a scope: you, everyone on this repository, or you in this repository only. Context7 works without a key on shared anonymous rate limits; for your own limits, create a key on the Context7 dashboard and set it as CONTEXT7_API_KEY before starting Claude Code.

A skill that calls Context7

Right after this, the video builds a second skill that looks up open-source documentation through Context7. The same idea works as a small skill of your own. The plugin's server is named context7, so its two tools are resolve-library-id and query-docs under the plugin's prefix:

---
name: library-docs
description: Look up current documentation for an open-source library with Context7. Use when asked how to use a library or what its latest API is.
allowed-tools: mcp__plugin_context7_context7__resolve-library-id mcp__plugin_context7_context7__query-docs
---

1. Find the library named in $ARGUMENTS with resolve-library-id.
2. Query its docs for the question with query-docs.
3. Answer from what the docs returned, with one short code example.
4. Say which library version the docs describe.

/library-docs langchain how do I create a deep agent? runs it. $ARGUMENTS becomes everything typed after the skill name, and allowed-tools lets the two Context7 calls run without a prompt for that turn.

Related

• Previous: MCP
• Next: Routines and scheduled tasks
• Reference: Claude Code docs