Skip to main content
Tips

practitioner tip

The 6 Claude Code Extension Points, Untangled

AIErudit angle: Skills, Hooks, MCP, Sub-agents, Agent Teams, and Plugins solve genuinely different problems, and picking the wrong one is a common source of over-engineered AI setups. This mental model — each mechanism answers one question (what to do, what to access, who works, when to automate, how to share) — lets you reach for the right extension point on purpose.

extension-pointsmcpclaude-codetips
Tip visual

Published: Feb 10, 2026 Updated: 2026-05-25

Original: https://wmedia.es/en/tips/claude-code-skills-hooks-mcp-plugins-comparison

Claude Code offers six different ways to extend its behavior: Skills, Hooks, MCP, Sub-agents, Agent Teams, and Plugins. They solve different problems, but their boundaries blur quickly. This tip gives you a mental model to tell them apart.

TL;DR Most people only need Skills and MCP. Add hooks when you want automation, sub-agents when context window pressure hits, and plugins when you need to share. Agent Teams are experimental — explore them when sub-agents aren't enough.

The core distinction: each one answers a different question.

QuestionMechanism
What should Claude do?Skills
What can Claude access?MCP
Who does the work?Sub-agents
Who does the work, collaboratively?Agent Teams
When should something happen automatically?Hooks
How do you package and share all of the above?Plugins

The 6 mechanisms

1. Skills — What Claude should do

Markdown files with instructions that Claude follows as slash commands or auto-activates when relevant. Reusable prompts with structure: arguments, frontmatter, subagent forking.

~/.claude/skills/review/SKILL.md → /review src/App.tsx

If you want to build a skill that actually gets invoked and doesn't sit unused, there are 5 rules Anthropic publishes ("pushy" description, SKILL.md <500 lines, evals against a baseline, etc.) — the short guide before you touch your first SKILL.md.

2. MCP — What Claude can access

Model Context Protocol connects Claude to external tools: databases, APIs, GitHub, Notion. MCP doesn't tell Claude what to do — it gives Claude hands to reach beyond its sandbox. When the time comes to install your own, start with the five MCP servers that earn the slot before bolting on the rest.

claude mcp add --transport http notion https://mcp.notion.com/mcp

3. Sub-agents — Who does the work

Separate Claude instances that handle tasks in isolation. Each gets its own context window, tools, and permissions. They work independently and report results back.

"Use the code-reviewer subagent to check the auth module"
→ Subagent works independently → Returns summary

4. Agent Teams — Who does the work, together (experimental)

Multiple Claude instances working as a coordinated team. Unlike sub-agents, teammates message each other, share a task list, and challenge each other's findings.

"Create a team: one on security, one on performance, one on tests"
→ 3 independent sessions → Cross-team discussion → Synthesis

We dive into the flag to enable them, the 4 use cases Anthropic recommends, and the limitations in the dedicated Agent Teams tip.

5. Hooks — When things happen automatically

Shell commands triggered at specific lifecycle points: before a tool runs, after a file is edited, when a session starts. No AI involved — pure automation.

{ "PostToolUse": [{ "matcher": "Write", "hooks": [{ "type": "command", "command": "npm run lint" }] }] }

6. Plugins — How you share everything

Distributable packages that bundle skills, agents, hooks, and MCP servers into an installable unit. Think npm packages for Claude Code extensions.

my-plugin/
├── .claude-plugin/plugin.json
├── skills/        ← reusable skills
├── agents/        ← custom subagents
├── hooks/         ← lifecycle automation
└── .mcp.json      ← external tool connections

Comparison

MechanismQuestion it answersDefined inAI involved?
SkillsWhat to doSKILL.md (markdown)Yes — Claude follows them
MCPWhat to access.mcp.json / CLI configNo — protocol bridge
Sub-agentsWho works.claude/agents/ (markdown)Yes — separate instance
Agent TeamsWho collaboratesNatural languageYes — multiple instances
HooksWhen to automatesettings.json (JSON)No — shell commands
PluginsHow to distribute.claude-plugin/ (package)No — packaging format

For a deep dive into Skills with advanced patterns, see the full Skills article.

Related: When features load into context

Official docs: Skills · MCP · Sub-agents · Agent Teams · Hooks · Plugins