Claude Code Plugin

[ncrmro]

Current State vs Plugin Architecture

DeepWork currently uses:

• CLI tool (deepwork install/sync) that generates platform-specific files
• Adapter pattern to support multiple platforms (Claude, Gemini)
• Jinja2 templates generating skills into .claude/skills/
• Hooks synced to .claude/settings.json

Claude Code plugins provide:

• Single distributable package with versioning
• Marketplace distribution (GitHub repos, custom marketplaces)
• Namespaced skills (/deepwork:job-name)
• Bundled hooks, MCP servers, LSP servers

Deployment Options

Option 1: Plugin Wrapper (Recommended)

Create a Claude Code plugin that wraps DeepWork:

deepwork-plugin/
├── .claude-plugin/
│ └── plugin.json # Plugin manifest
├── skills/
│ ├── install/SKILL.md # /deepwork:install
│ ├── sync/SKILL.md # /deepwork:sync
│ └── / # From deepwork sync
├── hooks/
│ └── hooks.json # rules_check hook
├── .mcp.json # Optional MCP server
└── scripts/
└── deepwork_wrapper.sh # Shell hooks

Pros:

• Single install command: /plugin install deepwork@marketplace
• Versioned releases via GitHub
• Works alongside existing adapter system
• Users don't need Python/pipx installed

Cons:

• Need to bundle Python runtime or assume it's installed
• Two-step process: install plugin + run deepwork install