Claude Code Plugins: What They Are, How They Differ From Skills and MCP, and Which to Install

A Claude Code plugin is an installable bundle that can package slash commands, subagents, skills, hooks, and MCP servers into a single unit, distributed through plugin marketplaces and installed with the /plugin command. Skills change what Claude knows, MCP servers change what it can connect to, and a plugin is the packaging format that can ship both at once, plus the commands and hooks that wire them together. The Skillselion catalog tracks 2,385 plugins alongside 60,665 skills and 8,427 MCP servers (Skillselion catalog, July 2026); this guide covers what plugins actually are, how the three types differ, and which plugins are worth installing.

Key takeaways

• A plugin is a distribution format, not a capability type. One install can add commands, subagents, skills, hooks, and MCP servers in a single step.
• Plugins ship through marketplaces: git repos with a .claude-plugin/marketplace.json manifest. Add one with /plugin marketplace add owner/repo, then /plugin install name@marketplace.
• The plugin/skill/MCP confusion is real. The Skillselion catalog re-typed 4,788 listings that marketplaces had published under the wrong type (Skillselion catalog cleanup, July 2026).
• Use a bare skill when you only need knowledge or a procedure. Reach for a plugin when the workflow needs hooks, commands, or an MCP server wired together.
• The most-starred plugins solve harness problems, not domain problems: memory, workflow discipline, and session visibility.

What is a Claude Code plugin?

A plugin is a directory with a .claude-plugin/plugin.json manifest that Claude Code can install as one unit. Inside, it can carry any mix of five component types:

• Slash commands - custom /commands that trigger a prompt or workflow.
• Subagents - specialized agents Claude can delegate to (a reviewer, a migration writer).
• Skills - SKILL.md instruction folders Claude loads when a task matches.
• Hooks - scripts that fire on lifecycle events, like after every file edit or before a commit.
• MCP servers - external tool connections declared in the plugin, so installing the plugin also configures the server.

Distribution runs through marketplaces: any git repo with a .claude-plugin/marketplace.json listing its plugins. There is no single central store, which is why directories exist to make the ecosystem searchable. The Skillselion marketplaces hub tracks 10,264 marketplace repos (Skillselion catalog, July 2026).

What is the difference between plugins, skills, and MCP servers?

The three get conflated constantly, including by marketplace publishers themselves: the Skillselion catalog cleanup re-typed 4,788 listings that marketplaces had shipped mislabeled, mostly plugins filed as skills (Skillselion catalog cleanup, July 2026). The distinction is simplest as a question each one answers:

• Skill - "what should Claude know or do for this kind of task?" A folder of markdown instructions plus optional scripts. No install ceremony beyond copying it into .claude/skills/. See what Claude Code skills are for the full picture.
• MCP server - "what external system should Claude be able to call?" A running process exposing tools over the Model Context Protocol: databases, browsers, APIs. See what MCP is.
• Plugin - "how do I ship a working setup of the above, plus commands and hooks, as one install?" It is the container, not the contents.

A plugin containing a single skill behaves almost identically to that skill installed bare. The differences show up at the edges: versioned updates through the marketplace, team-wide install via a shared marketplace config, and the ability to add hooks or an MCP server later without changing the install story. For a deeper head-to-head on the other pair, see Claude skills vs MCP.

How do you install a Claude Code plugin?

Two steps inside a Claude Code session:

1. Add the marketplace that hosts the plugin:

/plugin marketplace add anthropics/claude-code

Any owner/repo with a marketplace manifest works, as do full git URLs.

2. Install the plugin from that marketplace:

/plugin install plugin-name@marketplace-name

Running /plugin with no arguments opens an interactive menu where you can browse marketplaces, install, disable, or uninstall without memorizing names. The same verbs exist on the CLI (claude plugin install ...) for scripting and dotfiles. Installed plugins land in your user scope by default; a repo can also pin plugins for every collaborator via .claude/settings.json, which is the cleanest way to standardize a team. If the target turns out to be a bare skill rather than a plugin, the flow differs; see how to install a Claude skill.

When does a plugin beat a bare skill?

Skills win on simplicity: one folder, no marketplace, works across Claude Code, Claude Cowork, and the Agent SDK. Default to a skill when the job is knowledge, conventions, or a repeatable procedure.

A plugin earns its extra machinery when any of these hold:

• The workflow needs hooks. A formatter that runs after every edit or a check before each commit cannot be expressed as a skill; skills only advise the model, hooks execute deterministically.
• You want commands, not triggers. Skills activate when Claude judges them relevant; a slash command fires exactly when you type it.
• An MCP server is part of the deal. Bundling the server config means one install instead of instructions plus manual MCP setup.
• You maintain it for a team. Marketplace versioning beats "re-copy the folder when it changes."

Which Claude Code plugins are worth installing?

Ranked by GitHub stars in the Skillselion catalog (July 2026 snapshot), skipping docs collections and framework repos that are not installable plugins:

1. [claude-mem](/plugin/thedotmack/claude-mem) - 86,781 stars. Persistent memory: hooks capture each session, compress it, and surface relevant context when you return. The canonical example of a plugin doing what a skill cannot. 2. [get-shit-done](/plugin/gsd-build/get-shit-done) - 64,734 stars. Spec-driven development workflow: define what you are building, engineer context, then execute, instead of improvising every prompt. 3. [wshobson/agents](/plugin/wshobson/agents) - 37,767 stars. A collection of 75 plugins exposing narrow subagents and workflows across frontend, backend, DevOps, testing, and security. 4. [claude-hud](/plugin/jarrodwatts/claude-hud) - 26,318 stars. A statusline HUD showing context usage, active tools, and running agents, so long sessions stop being opaque. 5. [knowledge-work-plugins](/plugin/anthropics/knowledge-work-plugins) - 22,524 stars. Anthropic's own open-source plugin repo, aimed at knowledge work in Claude Cowork; useful as a reference implementation of plugin structure.

Browse the full ranked set on the plugin hub.

Install counts come from the skills.sh registry via the Skillselion catalog, snapshot July 2026; the skills.sh registry counts skill installs only, so plugin popularity above is reported in GitHub stars from the same catalog snapshot.

Common questions