claude-plugin-creator

Claude Plugin Creator

Guide for creating Claude Code plugins. For full reference, see references/plugins-reference.md.

Step 1: Decide What to Build

Ask the user these questions to determine the right approach:

Q1: Will this be used in one project or shared across projects/team?

• One project → standalone (.claude/ directory) or skill
• Shared → plugin

Q2: Should Claude use it automatically or should the user invoke it explicitly?

• Automatically (context-based) → skill (SKILL.md)
• Explicitly (/command) → command (markdown file)

Q3: Does it need to connect to external services/APIs?

• Yes → MCP server (standalone or inside plugin)
• No → skill or command

Q4: Should it run code automatically on events (file save, tool use, etc.)?

• Yes → hook
• No → skill or command

Q5: Does it need a specialized AI agent with its own system prompt?

• Yes → agent
• No → skill or command

Decision Summary

Need	Solution	Location
Personal, one project, auto-triggered	Standalone skill	.claude/skills/<name>/SKILL.md
Personal, one project, user-invoked	Standalone command	.claude/commands/<name>.md
Shared, auto-triggered	Plugin with skill	plugin/skills/<name>/SKILL.md
Shared, user-invoked	Plugin with command	plugin/commands/<name>.md
External API access	MCP server	.mcp.json (standalone or plugin)
Event automation	Hook	hooks/hooks.json (standalone or plugin)
Specialized AI role	Agent	agents/<name>.md
Multiple of the above	Plugin combining components	See plugin structure below

Scope: Who Can Access It?

Scope	File	Visibility
user	~/.claude/settings.json	You only, all projects
project	.claude/settings.json	Team via git, this project
local	.claude/settings.local.json	You only, this project (gitignored)
managed	Admin-controlled	Org-wide, read-only

Step 2: Create the Plugin

Directory Structure

my-plugin/
├── .claude-plugin/
│   └── plugin.json           # Only this file goes here
├── skills/                   # Auto-triggered by context
│   └── skill-name/
│       └── SKILL.md
├── commands/                 # User-invoked via /plugin:command
│   └── command-name.md
├── agents/                   # Specialized subagents
│   └── agent-name.md
├── hooks/
│   └── hooks.json            # Event handlers
├── bin/                      # Executables added to PATH
├── settings.json             # Default settings
├── .mcp.json                 # MCP server configs
└── .lsp.json                 # LSP server configs

Critical rules:

• Components go at plugin root, NOT inside .claude-plugin/
• Plugin name must be kebab-case
• Skills must use SKILL.md (exact case)

plugin.json (minimal)

{
  "name": "my-plugin",
  "description": "What the plugin does",
  "version": "1.0.0"
}

plugin.json (full)

{
  "name": "my-plugin",
  "version": "1.0.0",
  "description": "What the plugin does",
  "author": { "name": "Name", "email": "[email protected]" },
  "homepage": "https://docs.example.com",
  "repository": "https://github.com/user/plugin",
  "license": "MIT",
  "keywords": ["keyword1"],
  "userConfig": {
    "api_key": { "description": "API key", "sensitive": true }
  }
}

User config values available as ${user_config.KEY} in configs and CLAUDE_PLUGIN_OPTION_<KEY> env vars.

Skill (SKILL.md)

---
name: skill-name
description: What it does. Use when [trigger phrases].
---

Instructions for Claude.

Command (commands/name.md)

---
description: What it does
disable-model-invocation: true
---

Instructions. User input: $ARGUMENTS

Agent (agents/name.md)

---
name: agent-name
description: What it specializes in
model: sonnet
maxTurns: 20
---

System prompt here.

Hook (hooks/hooks.json)

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Write|Edit",
      "hooks": [{ "type": "command", "command": "${CLAUDE_PLUGIN_ROOT}/scripts/lint.sh" }]
    }]
  }
}

MCP Server (.mcp.json)

{
  "mcpServers": {
    "server-name": {
      "command": "${CLAUDE_PLUGIN_ROOT}/servers/server",
      "args": ["--config", "${CLAUDE_PLUGIN_ROOT}/config.json"]
    }
  }
}

Step 3: Test

claude --plugin-dir ./my-plugin

Inside session: /reload-plugins after changes. Validate: claude plugin validate .

Step 4: Distribute

Option A: Marketplace (for team/community)

Create .claude-plugin/marketplace.json in marketplace repo:

{
  "name": "marketplace-name",
  "owner": { "name": "Team Name" },
  "plugins": [
    {
      "name": "plugin-name",
      "source": "./plugins/plugin-name",
      "description": "What it does",
      "version": "1.0.0"
    }
  ]
}

Plugin sources can be: relative path ("./plugins/x"), GitHub ({ "source": "github", "repo": "owner/repo" }), git URL, git subdirectory, or npm package.

Host on GitHub, add with: /plugin marketplace add owner/repo

Option B: Direct sharing

Host plugin directory on GitHub. Users install with --plugin-dir or add to their .claude/ config.

Option C: Team auto-install

Add to project .claude/settings.json:

{
  "extraKnownMarketplaces": {
    "team-tools": { "source": { "source": "github", "repo": "org/plugins" } }
  },
  "enabledPlugins": { "my-plugin@team-tools": true }
}

Option D: Official marketplace

Submit at claude.ai/settings/plugins/submit or platform.claude.com/plugins/submit.

Converting Standalone to Plugin

1. mkdir -p my-plugin/.claude-plugin
2. Create plugin.json with name
3. Copy .claude/commands/ → my-plugin/commands/
4. Copy .claude/skills/ → my-plugin/skills/
5. Copy .claude/agents/ → my-plugin/agents/
6. Move hooks from settings.json → my-plugin/hooks/hooks.json
7. Test: claude --plugin-dir ./my-plugin

Environment Variables in Plugins

• ${CLAUDE_PLUGIN_ROOT} — plugin install path (changes on update)
• ${CLAUDE_PLUGIN_DATA} — persistent data path (survives updates)