05 — Plugins

A plugin bundles agents, skills, slash commands, and hooks around a theme — PR review, spec-driven dev, TDD. Install one and get a whole workflow out of the box instead of wiring each piece yourself.

Plugins vs MCPs

These two extension points get confused a lot. The distinction:

MCPs expose tools Claude can call — they're about connecting Claude to external services or data.
Plugins extend Claude Code's own behavior — new commands, new agents, new ways of working inside the tool itself.

An MCP gives Claude access to your GitHub repo. A plugin gives Claude a /pr-review command that knows how to structure a code review and which agents to call. They're complementary, not alternatives.

The marketplace

Plugins live in the official registry: anthropics/claude-plugins-official. Browse and search via the CLI:

claude plugins search          # browse available plugins
claude plugins info superpowers  # details on a specific plugin

Install, update, and list:

claude install superpowers     # install
claude update superpowers      # update to latest
claude plugins list            # see what's installed

All plugins install at user scope — ~/.claude/plugins/ — so they're available across all your projects.

What plugins actually give you

A plugin can contain any combination of:

Skills — reusable workflows Claude knows how to run (e.g. "systematic-debugging", "writing-plans")
Agents — specialist sub-agents Claude can delegate to (e.g. a dedicated PR reviewer)
Slash commands — /commands that trigger specific workflows
Hooks — code that runs automatically on file writes, session start, or commit
Bundled MCPs — the plugin installs and wires an MCP server for you

Plugins worth trying

Pick 1-2 to start; evaluate before adding more.

superpowers (by Obra) — battle-tested skills for planning, TDD, brainstorming, systematic debugging, and spec-driven development. The highest-ROI plugin available: each skill is a structured prompt sequence with clear inputs and outputs. Start here.

claude-mem — persists context across sessions so Claude picks up where it left off, not from a cold start. Exposes a mem-search skill and a /resume command. Worth it if you find yourself re-explaining your stack every session.

hookify — turns natural-language feedback ("stop doing X") into a saved hook automatically. Faster than writing hooks by hand.

frontend-design — design-review skills and UI pattern guidance. Useful if you're building UI and want a structured design pass alongside code generation.

All four are publicly distributed via claude install <name>. No private infra, no keys needed beyond what the plugin itself documents.

Install one vs build your own

Install from the marketplace when the workflow it bundles matches how you already want to work. If a plugin's opinionated flow doesn't fit, you're better off building a lightweight version yourself — a skill is just a markdown file in ~/.claude/skills/, a hook is a script in ~/.claude/hooks/.

Building your own forces you to understand the pieces. Installing a plugin is faster but you inherit its assumptions. Both are valid; just know which one you're doing.

Evaluating a plugin before you commit

Check three things: what commands and skills it adds (do they overlap with something you already have?), whether it installs any MCPs (do you trust those?), and whether the hooks it registers run on events you care about. claude plugins info <name> surfaces all of this before you install.