Interactive map

Claude Code Ecosystem

All building blocks and how they work together — click any block for details

Agent Teams

Multiple sessions coordinating in parallel — the highest level

Session 1

CLAUDE.mdSkillsHooksSubagentsMCP

Session 2

CLAUDE.mdSkillsHooksSubagentsMCP

Session 3

CLAUDE.mdSkillsHooksSubagentsMCP

each session = all of this ↓

Single Session — one Claude Code instance (Terminal / VSCode)

Context Layer — CLAUDE.md

The foundation. Always loaded into every conversation.

CLAUDE.md

Team rules — shared via repo

CLAUDE.local.md

Personal tweaks — not shared

~/.claude/CLAUDE.md

Global — all your projects

feeds context to everything below ↓

Skills

On-demand knowledge — loads when semantically matched

Personal — ~/.claude/skills/

Project — .claude/skills/ — shared via repo

Enterprise — Managed, highest priority

SKILL.md < 500 linesRequest-driven

Hooks

Scripts that fire automatically on system events

PreToolUse — Before actions — validate, block

PostToolUse — After actions — format, lint

Stop — When Claude finishes responding

Notification — On alerts, custom triggers

exit(0) proceedexit(2) block

Subagents

Delegated workers with isolated context

Explorer — Built-in, read-only codebase search

Plan — Built-in, research + planning

Custom — .claude/agents/ — your specialists

Don't see skills unless listed

Slash Commands

Manual triggers — you type them, unlike skills

/review — Predefined review prompt

/commit — Commit message generation

Custom — Your own slash commands

Manual ≠ Skills (automatic)

Plugins

Installed packages — bundle skills + commands + subagents

Bundled Skills — Task-specific knowledge

Bundled Commands — Custom slash commands

Bundled Subagents — Specialized workers

Distributable via marketplacesLowest skill priority

Session Controls

Manage context window and behavior

Plan Mode — Read-only — Shift+Tab

Extended Thinking — Deeper — Option+T

Escape / Double Escape — Interrupt / Rewind

/compact · /clear — Summarize / Full reset

↑ sends requests — returns data ↓

MCP Servers

External tools — add capabilities, not knowledge

Figma — Design context, screenshots, metadata

Notion — Pages, databases, search

Playwright — Browser automation, testing

GitHub — Repos, PRs, issues

claude mcp add --scope userScopes: local / project / user

Skills → request-driven (automatic)

Hooks → event-driven (always fire)

CLAUDE.md → always loaded

MCP → tools, not knowledge

Priority: Enterprise > Personal > Project > Plugin

Based on Anthropic Academy + Claude Code docs — by Kike López