Plugins

"Plugin" is overloaded. In the AI-agent world, a plugin is a packaged bundle that adds features, commands, skills, MCP servers, hooks, to a harness like Claude Code.

What a plugin contains

A plugin (in the Claude Code sense) is a directory with metadata declaring what it adds. A single plugin can bundle:

• One or more MCP servers (tools the agent can use)
• One or more skills (named slash-commands with instructions)
• One or more hooks (code that runs on events)
• Shared config or assets

Why plugins exist

Individually, MCPs and skills are great. But connecting them is glue. A plugin bundles the pieces into a one-click install. Example: a "Telegram channel" plugin might ship an MCP server (for reading/writing messages), a skill (for configuring the bot), and a hook (to notify on certain events).

Plugin vs MCP vs Skill

Short answer:

• MCP, adds tools/data the agent can use
• Skill, adds a named prompt/instruction the agent can invoke
• Plugin, a package that bundles any combination of the above

See Plugins vs MCP for the deeper comparison.

How plugins are distributed

Most plugins today are distributed through marketplaces. Anthropic maintains an official marketplace of first-party plugins. There's a growing ecosystem of community plugins on GitHub. See Marketplace.

Installing a plugin

In Claude Code:

/plugin install telegram@claude-plugins-official

The plugin is downloaded, any MCP servers are registered, any skills become invokable, any hooks are wired up. Removing is one command.

When to reach for a plugin vs, just an MCP

• Just an MCP: you want tools. Nothing more.
• Plugin: you want tools + specific workflows + event handling + config, all as one unit.

For most users, the split is: MCPs are the primitives; plugins are the products.