Plugin Anatomy — Claude Code Plugin Structure

A Claude Code plugin is a directory installed via --plugin-dir that extends Claude Code with commands, skills, agents, hooks, and MCP integrations.

Canonical Directory Layout

plugins/<plugin-name>/
├── .claude-plugin/
│   └── plugin.json          # Manifest (required)
├── commands/
│   └── <command-name>.md    # Slash commands (/plugin-name:command-name)
├── skills/
│   └── <skill-name>/
│       └── SKILL.md         # Skill definitions (triggered by Claude)
├── agents/
│   └── <agent-name>.md      # Autonomous subagents
├── hooks/
│   ├── hooks.json           # Hook event configuration
│   └── scripts/             # Bash scripts called by command hooks
├── mcp-template.json        # MCP server configuration template (optional)
├── scripts/                 # Plugin utility scripts (optional)
└── README.md                # Required for marketplace listing

plugin.json — Manifest

Required fields: name, version, description, author.

{
  "name": "plugin-name",
  "version": "0.1.0",
  "description": "What this plugin does",
  "author": {
    "name": "Author Name",
    "email": "author@example.com"
  }
}

Naming rules: kebab-case, lowercase only, no spaces.

Commands (commands/*.md)

Slash commands invoked as /plugin-name:command-name.

Required frontmatter field: description.

Optional frontmatter fields:

• argument-hint — shown in autocomplete (e.g., "[path]")
• allowed-tools — tools the command may invoke (e.g., Bash, Read)
• model — override model (haiku, sonnet, opus, or inherit)

---
description: Short description shown in autocomplete
argument-hint: "[optional-arg]"
allowed-tools: Bash, Read
---

# Command Title

Command body in Markdown. This is the prompt/instructions for Claude.

Skills (skills/<name>/SKILL.md)

Skills are passive knowledge loaded when Claude detects the trigger described in description. They are NOT slash commands.

Required frontmatter fields: name, description.

---
name: skill-name
description: >
  Trigger conditions — what user actions or keywords activate this skill.
  Be specific so Claude knows when to load it.
---

# Skill Content

The actual knowledge, guidelines, or instructions Claude should follow.

Agents (agents/*.md)

Autonomous subagents spawned by Claude via the Agent tool. They receive a focused context and available tools defined in their frontmatter.

Required frontmatter fields: name, description, model, color.

---
name: agent-name
description: Purpose and when to spawn this agent
model: claude-haiku-4-5-20251001
color: blue
tools: Read, Grep, Glob
---

# Agent System Prompt

Instructions for the agent's behaviour and focus.

Hooks (hooks/hooks.json)

Event-driven automation. Supported events:

Event	When
SessionStart	Claude Code session begins
SessionEnd	Session ends
PreToolUse	Before any tool executes
PostToolUse	After any tool executes
UserPromptSubmit	User submits a message
Stop	Claude finishes a response
SubagentStop	A subagent finishes
Notification	System notification
PreCompact	Before context compaction

Hook types:

• command — runs a bash script; supports $CLAUDE_PLUGIN_ROOT
• prompt — LLM-driven decision; supported on Stop, SubagentStop, UserPromptSubmit, PreToolUse

{
  "SessionStart": [
    {
      "matcher": ".*",
      "hooks": [
        {
          "type": "command",
          "command": "bash ${CLAUDE_PLUGIN_ROOT}/hooks/scripts/on-start.sh",
          "timeout": 10,
          "description": "Run setup on session start"
        }
      ]
    }
  ]
}

MCP Integration (mcp-template.json)

Template for connecting external MCP servers. Users copy and fill in credentials.

{
  "mcpServers": {
    "server-name": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@scope/mcp-server"],
      "env": {
        "API_KEY": "${YOUR_API_KEY}"
      }
    }
  }
}

Plugin Studio Component Types

Plugin Studio uses component-type awareness to provide contextual editing:

Directory pattern	Component type	Frontmatter
commands/*.md	Command	description (required), argument-hint, allowed-tools, model
skills/*/SKILL.md	Skill	name, description (both required)
agents/*.md	Agent	name, description, model, color (all required)
hooks/hooks.json	Hooks config	JSON (validated by plugin-dev)
.claude-plugin/plugin.json	Manifest	JSON (validated by plugin-dev)
mcp-template.json	MCP config	JSON