Live on Product Hunt Intraface launched today. Less steering, more vacations. See the launch
Start for free

What are Claude Code plugins?

A Claude Code plugin is a directory of skills, subagents, hooks and MCP servers that Claude Code installs and loads as one unit. A small manifest at .claude-plugin/plugin.json gives it a name, which becomes the prefix on its skills. Most plugins come from a marketplace: a catalog file that lists plugins and says where to fetch each one.

Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.

How it works

A marketplace lists the plugin, Claude Code installs the directory as a unit, and each component inside it becomes part of your sessions.a catalog, not a storemarketplacethe pluginone directory, a manifestClaude Codeyour sessionsskills/plugin-name:skillhooksrun on eventssubagents,MCP servers1. lists, points to2. installed as a unit3. adds3. adds3. adds
  1. Someone writes a plugin: a directory with .claude-plugin/plugin.json and, as needed, skills/, agents/, hooks/hooks.json and .mcp.json.

  2. They list it in a marketplace: a repository whose .claude-plugin/marketplace.json says where each plugin lives. Anthropic's official one is added for you.

  3. You run /plugin, pick it from the Discover tab and choose a scope: you everywhere, everyone in this project, or you in this project only.

  4. Claude Code fetches it to ~/.claude/plugins/ and loads every component at startup. Its skills appear as /plugin-name:skill-name.

  5. Disable it in /plugin or with claude plugin disable: its descriptions leave your context and its hooks and servers stop.

Source: Claude Code docs: plugins overview

Plugins versus standalone skills and hooks

Nothing in a plugin is unique to plugins. A skill in ~/.claude/skills/ works in every project on your machine, a hook in .claude/settings.json fires for everyone in the repository, and an MCP server in .mcp.json connects without any wrapper. The plugin is a packaging layer: the same components, moved under one directory and given a name, so that they install together, update together and can be turned off together.

Two things change when you package. Location: the files move to the plugin root as skills/, agents/, hooks/hooks.json and .mcp.json. Naming: skills and subagents gain the plugin name as a prefix, so /review becomes /my-plugin:review and two plugins can each ship a review without colliding. Hooks get no prefix; they are matched by event, not by name.

Example: the layout of a plugin

my-plugin/
├── .claude-plugin/plugin.json   # the manifest: name, description, version, author
├── skills/review/SKILL.md       # becomes /my-plugin:review
├── agents/reviewer.md           # a subagent Claude can delegate to
├── hooks/hooks.json             # same shape as "hooks" in a settings file
└── .mcp.json                    # MCP servers to connect while enabled

Only name is required in the manifest, and only plugin.json belongs inside .claude-plugin/; components saved there don't load. Every directory is optional, so a plugin with one skill is valid. While you build it, claude plugin validate ./my-plugin checks the manifest and every frontmatter, and claude --plugin-dir ./my-plugin loads it for one session with no marketplace and nothing written to your settings.

When to make one

Make a plugin when a setup should outlive one project or one person: a team's review skill, a set of safety hooks, and the MCP server they rely on, installed by everyone with one command and updated from one place. Keep things standalone while they serve one repository or only you. Anthropic's plugin creation docs say the same: convert when you want to share, install in several projects, or publish versioned releases.

An enabled plugin is part of every session, not only the ones that use it. Each skill, agent and command that Claude can invoke on its own adds its name and description to context on every turn, and the plugin's MCP servers and hooks run wherever it is enabled. That is the trade: convenience for a standing cost, and a plugin that runs as you with your permissions.

Common mistakes

  • Mixing up the two marketplaces. A plugin marketplace is a marketplace.json in a repository you add with /plugin marketplace add. Claude Marketplace is the website at claude.com where you browse; it isn't something you add.
  • Pointing --plugin-dir at a marketplace root. It takes one plugin's folder. Point it at a marketplace and nothing loads and nothing complains.
  • Installing at the wrong scope. Project scope writes the plugin into the committed .claude/settings.json, so everyone in the repository gets it enabled, though each person still installs it on their own machine.
  • Expecting plugins in cloud sessions. A session at claude.ai/code doesn't load the plugins in your local settings.
  • Trusting a plugin because of where it came from. Official, community and third-party marketplaces are tiers of naming, not of safety. Whatever the tier, read the hooks and MCP config before you enable them.

FAQ

Do I need a Claude Code plugin to use a skill?
No. Skills, subagents, hooks and MCP servers all work on their own from your project's .claude/ directory or your home directory. A plugin is for packaging several of them so teammates get the set with one command and receive updates from its marketplace.
Where do Claude Code plugins come from?
From marketplaces, which are repositories with a .claude-plugin/marketplace.json listing plugins. Claude Code adds Anthropic's official one, claude-plugins-official, the first time you start an interactive session. Your team or any third party can publish its own, and you can load a folder directly with --plugin-dir.
What do Claude Code plugins cost while enabled?
The name and description of every skill, agent and command Claude can invoke on its own sit in context on every turn, in every session, even when nothing from the plugin runs. Its MCP servers run beside each session and its hooks fire on their events. The official marketplace shows a context cost estimate before you install.

Stop sending HTMLover Slack.

Create an account in a minute, connect your coding agent, and make your first board.