Getting Started
Building Blocks
How plugins bundle commands, agents, MCPs, and skills into shareable packages — and how to install, manage, and build your own.
This module ties together everything you've learned about the five building blocks. Plugins are how they get packaged, shared, and distributed.
A plugin is a folder that bundles any combination of the five building blocks. Some plugins are nothing more than an MCP wrapper. Others combine commands, skills, agents, and hooks into sophisticated workflows.
| Plugin | MCPs | Commands | Skills | Agents | Hooks |
|---|---|---|---|---|---|
| context7 | ✓ | – | – | – | – |
| vercel | – | /deploy, /setup, /logs | deploy, setup, logs | – | – |
| feature-dev | – | /feature-dev | – | code-explorer, code-architect | – |
| security-guidance | – | – | – | – | PreToolUse (9 patterns) |
| compound-engineering | – | /plan, /review, /compound | plan, review | – | – |
What is a plugin in Claude Code?
The official Anthropic marketplace has 36+ curated plugins. Community marketplaces add hundreds more.
Install a plugin:
# Add a marketplace
/plugin marketplace add https://github.com/EveryInc/compound-engineering-plugin
# Install from it
/plugin install compound-engineeringNot all plugins are equal. Before installing, check:
allowed-tools in its commands — what tools does it pre-approve?You can always read a plugin's files after installation — they're just markdown and JSON on disk.
All plugin files are readable, editable markdown on disk. The override precedence is:
.claude/ in your repo) — highest priority~/.claude/) — middleYou can copy any plugin's agent, skill, or command into your own directories and customise it. Your version takes precedence.
What is the plugin override precedence when the same command exists at multiple levels?
The marketplace is a starting point, not a destination. The real power comes from forking plugins and customising them for your specific project, stack, and conventions.
A few plugins worth knowing about:
| Plugin | What It Does | Why It Matters |
|---|---|---|
| Compound Engineering (Every) | Plan → Work → Review → Codify loop | Formalises the compounding engineering habit from the Context module |
| Ralph Wiggum (Anthropic) | Persistent autonomous loops for long tasks | Lets Claude run for hours, fixing its own mistakes until done |
| Security Guidance | 9 PreToolUse hooks for security patterns | Passively catches secrets, injection flaws, insecure patterns |
| Context7 | Live documentation fetching via MCP | Avoids deprecated API usage in fast-moving frameworks |
The provides a programmatic interface to Claude Code — available as CLI, TypeScript, and Python libraries.
Integration into larger pipelines and workflows — CI/CD, automated code review, batch processing. If you just need Claude to review code in CI, claude --print "Review this PR" piped into a script might be all you need. The SDK is for more complex programmatic control.
import { createAgent } from "@anthropic-ai/claude-code";
const agent = createAgent({
system_prompt: "You are a code reviewer...",
allowTools: ["Read", "Glob", "Grep"],
model: "claude-sonnet-4-5-20250929",
});The SDK's system_prompt is the programmatic equivalent of an agent persona — the same concept from the Agents module, just written in code instead of a markdown file.
/install-github-app in Claude Code@Claude mentions in PRs and issues trigger Claude to respondThis is particularly powerful combined with the compounding engineering habit — team members can comment @claude add this to CLAUDE.md so you don't do it again during PR review, and Claude creates a PR updating the rules file.
claude --print "Review this PR" in a shell script covers many CI use cases. Don't over-engineer.Plugins are just bundles of building blocks — forkable and customisable. Once you understand the five pieces (MCPs, commands, skills, agents, hooks), you can read, evaluate, and modify any plugin. The marketplace gives you a starting point; your project's .claude/ directory is where the real value lives.