Understanding Claude Code's Full Stack: MCP, Skills, Subagents, and Hooks Explained

I used Claude Code for months as a glorified autocomplete. Quick edits. Boilerplate. The vibe coding tool.

Then I dug into MCP servers, slash commands, plugins, skills, hooks, subagents, and CLAUDE.md files. Claude Code is a framework for orchestrating AI agents, and most people use one or two features without seeing how they stack together. (It also runs in more places now: CLI, VS Code, JetBrains, a standalone desktop app, a web app at claude.ai/code, and iOS.)

This guide explains each concept in the order they build on each other, from external connections to automatic behaviors. (New to using LLMs for development? Start with my overview of how I use LLMs How I Use LLMs Learn how I use LLMs to improve my productivity and efficiency. aiproductivity May 25, 2025 for context.)

Claude Code is, with hindsight, poorly named. It’s not purely a coding tool: it’s a tool for general computer automation. Anything you can achieve by typing commands into a computer is something that can now be automated by Claude Code. It’s best described as a general agent. Skills make this a whole lot more obvious and explicit.

— Simon Willison, Claude Skills are awesome, maybe a bigger deal than MCP

The feature stack#

1. Model Context Protocol (MCP) — the foundation for connecting external tools and data sources
2. Claude Code core features — project memory, skills, subagents, and hooks
3. Plugins — shareable packages that bundle skills, hooks, and metadata
4. Skills — the unified extensibility model (replaces the old command/skill split)
5. Scheduled Tasks — cloud-based triggers that run Claude on a cron schedule

1) Model Context Protocol (MCP) — connecting external systems#

What it is. The Model Context Protocol What Is the Model Context Protocol (MCP)? How It Works Learn how MCP (Model Context Protocol) standardizes AI tool integration, enabling LLMs to interact with external services, databases, and APIs through a universal protocol similar to USB-C for AI applications. mcptypescriptai Aug 10, 2025 connects Claude Code to external tools and data sources. Think universal adapter for GitHub, databases, APIs, and other systems.

How it works. Connect an MCP server, get access to its tools, resources, and prompts as slash commands:

# Install a server
claude mcp add playwright npx @playwright/mcp@latest

# Use it
/mcp__playwright__create-test [args]

Remote servers. The MCP spec added streamable HTTP transport (replacing SSE) and OAuth 2.1 with PKCE for authentication. You can now connect to remote MCP servers without running a local stdio process.

The gotcha. MCP servers expose their own tools. They don’t inherit Claude’s Read, Write, or Bash unless you provide them.

Real-world example. Want to see MCP in action? Check out how to build an AI QA engineer with Playwright MCP Building an AI QA Engineer with Claude Code and Playwright MCP Learn how to build an automated QA engineer using Claude Code and Playwright MCP that tests your web app like a real user, runs on every pull request, and writes detailed bug reports. aitestingclaude-code +1 Dec 13, 2025 that tests your app like a real user.

2) Claude Code core features#

2.1) Project memory with CLAUDE.md#

What it is. Markdown files Claude loads at startup. They give Claude memory about your project’s conventions, architecture, and patterns.

How it works. Files merge hierarchically from enterprise → user (~/.claude/CLAUDE.md) → project (./CLAUDE.md). When you reference @components/Button.vue, Claude also reads CLAUDE.md from that directory and its parents.

Two additions since launch: CLAUDE.local.md files sit alongside CLAUDE.md but are gitignored, so you can add personal preferences without affecting the team. And ~/.claude/rules/ lets you split global instructions into separate files instead of one large ~/.claude/CLAUDE.md.

Example structure for a Vue app:

When you work on src/components/Button.vue, Claude loads context from:

1. Enterprise CLAUDE.md (if configured)
2. User ~/.claude/CLAUDE.md (personal preferences)
3. Project root CLAUDE.md (project-wide info)
4. src/components/CLAUDE.md (component-specific patterns)

What goes in. Common commands, coding standards, architectural patterns. Keep it concise — reference guide, not documentation. Need help creating your own? Check out this CLAUDE.md creation guide.

Here’s my blog’s CLAUDE.md:

# CLAUDE.md

## Project Overview

Alexander Opalic's personal blog built on AstroPaper - Astro-based blog theme with TypeScript, React, TailwindCSS.

**Tech Stack**: Astro 5, TypeScript, React, TailwindCSS, Shiki, FuseJS, Playwright

## Development Commands

```bash
npm run dev              # Build + Pagefind + dev server (localhost:4321)
npm run build            # Production build
npm run lint             # ESLint for .astro, .ts, .tsx
---
```

2.2) Skills — the unified extensibility model#

What they are. Folders with a SKILL.md file (plus optional helper scripts) that define reusable behaviors. Frontmatter controls invocation: Claude can auto-invoke based on task context, you can trigger manually with /skill-name, or both.

Key frontmatter fields:

Argument substitution: $ARGUMENTS for all args, $0, $1, $2 for positional, ${CLAUDE_SKILL_DIR} for the skill’s directory.

Key features:

• @file syntax to inline code
• allowed-tools: Bash(...) for pre-execution scripts

• XML-tagged prompts

  XML-Style Tagged Prompts: A Framework for Reliable AI Responses Learn how top AI engineers use XML-style prompts to consistently get structured, accurate responses from ChatGPT, Claude, and other LLMs. Step-by-step guide with real examples aillmprompt-engineering Dec 22, 2024 for reliable outputs

When to use. Repeatable workflows, domain expertise, automated enforcement. For a complete example of a git workflow built with skills, see my Slash Commands Guide How to Speed Up Your Claude Code Experience with Slash Commands Learn how to transform Claude Code from a chatbot into a deterministic engine using Slash Commands. This guide covers the technical setup and a complete 'Full Circle' workflow that automates your entire feature lifecycle. aiclaude-code Nov 22, 2025 . Want to create your own? Use this skill creation guide.

Example: a deploy skill that only you can trigger:

---
name: deploy
description: Deploy the application to production
disable-model-invocation: true
allowed-tools: Bash(npm:*), Bash(git:*)
---

Deploy to production:
1. Run tests
2. Build
3. Push to deployment target

This creates /deploy that only you can invoke — Claude cannot auto-trigger it.

Example: a skill that runs in a subagent:

---
name: deep-research
description: Research a topic thoroughly
context: fork
agent: Explore
allowed-tools: Read, Grep, Glob
---

Research $ARGUMENTS thoroughly. Find relevant files, read and analyze them, summarize findings.

When invoked, this spawns a separate context window. Your main conversation stays clean.

2.3) Subagents — specialized AI personalities for delegation#

What they are. Pre-configured AI personalities with specific expertise areas. Each subagent has its own system prompt, allowed tools, and separate context window. Claude delegates to them when a task matches their expertise.

Why use them. Keep your main conversation clean while offloading specialized work. Each subagent works in its own context window, preventing token bloat. Run multiple subagents in parallel for concurrent analysis.

Example structure:

---
name: security-auditor
description: Analyzes code for security vulnerabilities
tools: Read, Grep, Bash # Controls what this personality can access
model: sonnet # Optional: sonnet, opus, haiku, inherit
---

You are a security-focused code auditor.

Identify vulnerabilities (XSS, SQL injection, CSRF, etc.)
Check dependencies and packages
Verify auth/authorization
Review data validation

Provide severity levels: Critical, High, Medium, Low.
Focus on OWASP Top 10.

The system prompt shapes the subagent’s behavior. The description tells Claude when to delegate. The tools field restricts access.

Best practices: One expertise area per subagent. Minimal tool access. Use haiku for simple tasks, sonnet for complex analysis. Run independent work in parallel. Need a template? Check out this subagent creation guide.

Claude also ships three built-in subagent types you can use without writing a config file: Explore (fast, read-only, uses Haiku), Plan (research and architecture, read-only), and General-purpose (full tool access).

2.4) Hooks — automatic event-driven actions#

What they are. JSON-configured handlers in .claude/settings.json that trigger on lifecycle events. No manual invocation.

Available events (expanded from the original 7):

• Tool lifecycle: PreToolUse, PostToolUse
• Session lifecycle: SessionStart, Stop, SubagentStart, SubagentStop
• Task lifecycle: TaskCreated, TaskCompleted
• Environment: CwdChanged, FileChanged
• Permissions: PermissionDenied
• Context: PreCompact, PostCompact
• User input: UserPromptSubmit, Notification

Handler types:

• Command: Run shell commands (fast, predictable)
• Prompt: Let Claude decide with the LLM (flexible, context-aware)
• HTTP: POST JSON to an external service with auth headers
• Async: Add "async": true to run a hook in the background without blocking

PreToolUse hooks can also return updatedInput to modify tool arguments before execution.

Example: Auto-lint after file edits.

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/run-oxlint.sh"
          }
        ]
      }
    ]
  }
}

#!/usr/bin/env bash
file_path="$(jq -r '.tool_input.file_path // ""')"

if [[ "$file_path" =~ \.(js|jsx|ts|tsx|vue)$ ]]; then
  pnpm lint:fast
fi

Common uses: Auto-format after edits, require approval for bash commands, validate writes, initialize sessions. For a practical example, see how to set up desktop notifications Claude Code Notifications: Get Alerts When Tasks Finish (Hooks Setup) How to set up Claude Code notifications using hooks. Get desktop alerts when Claude finishes a task, needs your input, or requests permission, instead of watching the terminal. claude-codenotificationshooks +1 Nov 23, 2025 when Claude needs your attention. Want to build your own? Use this hook creation guide.

3) Plugins — shareable, packaged configurations#

What they are. Distributable bundles of skills, hooks, and metadata. Share your setup with teammates or install pre-built configurations.

Basic structure:

When to use. Share team configurations, package domain workflows Building My First Claude Code Plugin How I built a Claude Code plugin to generate skills, agents, commands, and more—and stopped copy-pasting boilerplate. claude-codeaitooling +1 Nov 8, 2025 , distribute opinionated patterns, install community tooling.

How it works. Install a plugin, get instant access. Hooks combine, skills appear in autocomplete and activate on context match. Ready to build your own? Check out this plugin creation guide.

4) Skills deep dive — where to put them and how they compose#

Skills were covered in section 2.2. This section focuses on discovery, placement, and composition patterns.

How Claude discovers skills. When you give Claude a task, it reviews available skill descriptions to find relevant ones. If a skill’s description field matches the task context, Claude loads the full instructions. You don’t need to invoke auto-invoked skills by name — Claude finds them.

Claude Skills are awesome, maybe a bigger deal than MCP

— Simon Willison, Claude Skills are awesome, maybe a bigger deal than MCP

Where to put them:

• ~/.claude/skills/ — personal, all projects
• .claude/skills/ — project-specific
• Inside plugins — distributable

Skills vs CLAUDE.md. Skills are modular chunks of a CLAUDE.md file. Instead of reviewing a massive document on every task, Claude loads specific skill instructions only when the task matches. Better context efficiency, same automatic behavior.

Controlling invocation with frontmatter. The old command/skill split is gone. A single SKILL.md file with different frontmatter covers every use case:

5) Scheduled Tasks and Triggers — running Claude without you#

What they are. Cloud-based cron jobs that run Claude Code on Anthropic’s infrastructure. You define a schedule and a prompt; Claude executes it on the cron interval, with full access to your project.

How to set one up. Use the /schedule command or claude trigger create from the CLI. Define the cron expression and the task prompt:

claude trigger create --schedule "0 9 * * 1" --prompt "Review open PRs and summarize status"

In-session polling. For shorter-lived recurring work, /loop 5m /your-command runs a slash command on an interval inside your current session.

Putting it all together#

How these features work together in practice:

1. Memory (CLAUDE.md) — Establish project context and conventions that Claude always knows
2. Skills — Define reusable behaviors, from manual workflows to auto-invoked expertise
3. Subagents — Offload parallel or isolated work to specialized agents
4. Hooks — Enforce rules and automate repetitive actions at key lifecycle events
5. Plugins — Package and distribute your entire setup to others
6. MCP — Connect external systems and make their capabilities available as commands
7. Scheduled Tasks — Run Claude on a cron schedule for recurring work

Example: A Task-Based Development Workflow#

A real-world workflow that combines multiple features:

Setup phase:

• CLAUDE.md contains implementation standards (“don’t commit until I approve”, “write tests first”)
• /load-context skill (disable-model-invocation: true) initializes new chats with project state
• update-documentation skill (auto-invoked) activates after implementations
• Hook triggers linting after file edits

Planning phase (Chat 1):

• Main agent plans bug fix or new feature
• Outputs detailed task file with approach

Implementation phase (Chat 2):

• Start fresh context with /load-context
• Feed in the plan from Chat 1
• Implementation subagent executes the plan
• update-documentation skill updates docs on its own
• /resolve-task skill marks task complete

Why this works: Main context stays focused on planning. Heavy implementation work happens in isolated context. Skills handle documentation. Hooks enforce quality standards.

Decision guide: choosing the right tool#

Feature comparison#

Real-world examples#