For months I’d been asking Claude to manually configure MCP servers for me, using the GitHub and Vercel CLIs directly, and recently asking Claude to write a custom skill for my content workflow. Then a friend mentioned /plugin and I typed it in and there’s a whole browser for installing stuff? I saw the list of skills Anthropic hosts and my brain basically melted. Plugins, skills, MCP, commands, hooks - what is with all these different tools? It’s hard to keep up.

Turns out Anthropic launched this in October 2025, but I completely missed it.

So I asked Claude to write me a post explaining it. Here’s what I learned.

Plugins Are the Container

Plugins are bundles you can install from /plugin. A plugin can contain any mix of:

  • MCP server configs (connections to external APIs)
  • Skills (specialized knowledge Claude loads automatically)
  • Hooks (automation rules that run before/after Claude acts)
  • Commands (saved prompts you trigger with /name)

Type /plugin to open the browser. It shows plugins from whatever marketplaces you have added - Anthropic has an official directory at claude.com/connectors with partner-built stuff from Figma, Vercel, Zapier, and others. There are also community marketplaces with hundreds more. You can add marketplaces with /plugin marketplace add.

WHAT’S A PLUGIN HERE In Claude Code, plugins are like app bundles. You install one thing and it might set up multiple capabilities - an MCP connection, some skills, maybe a command or two.

Skills Activate Automatically

Skills don’t work like the other stuff. You don’t invoke them. They activate automatically when Claude decides they’re relevant. I wrote about creating my first skill if you want to see how they work under the hood.

If you have a frontend-design skill installed and start building UI, Claude loads it without you doing anything. Ask Claude to create a PDF with the right skill installed, it just happens. Claude reads skill descriptions and matches them to what you’re doing.

THE ELI5 Skills are recipes (instructions on how to do something). MCPs are tools (the things that actually do actions, like connecting to a database or API). Skills tell Claude how, MCPs give Claude access.

This feels a little magical and also a little opaque. I can’t always tell when a skill is active vs when Claude is just doing its normal thing.

There’s some genuinely useful stuff in here. I really like context7 for fetching current library docs instead of relying on Claude’s training data. The discoverability is rough though - I had no idea any of this existed until someone told me.

That said, you don’t have to use /plugin if it feels like another thing to learn. You can always just ask Claude to set things up manually, which is what I was doing for months anyway. It worked fine.