MCP Design System Extractor

A Model Context Protocol (MCP) server that extracts component information from Storybook design systems. Connects to Storybook instances and extracts HTML, styles, and component metadata.

Demo

Installation

Using Claude CLI (Recommended)

claude mcp add design-system npx mcp-design-system-extractor@latest \
  --env STORYBOOK_URL=http://localhost:6006

With self-signed certificate:

claude mcp add design-system npx mcp-design-system-extractor@latest \
  --env STORYBOOK_URL=https://my-storybook.example.com \
  --env NODE_TLS_REJECT_UNAUTHORIZED=0

Using npm

npm install -g mcp-design-system-extractor

Then configure in your MCP client (see Environment Variables).

From Source

git clone https://github.com/freema/mcp-design-system-extractor.git
cd mcp-design-system-extractor
npm install && npm run build
npm run setup  # Interactive setup for Claude Desktop

Key Dependencies

• Puppeteer: Uses headless Chrome for dynamic JavaScript component rendering
• Chrome/Chromium: Required for Puppeteer (automatically handled in Docker)
• Works with built Storybook distributions

Features

• List Components: Get all available components from your Storybook with compact mode
• Extract HTML: Get the rendered HTML of any component (async or sync mode)
• Search Components: Find components by name, title, category, or purpose
• Component Dependencies: Analyze which components are used within other components
• Theme Information: Extract design system theme (colors, spacing, typography)
• External CSS Analysis: Fetch and analyze CSS files to extract design tokens
• Async Job Queue: Long-running operations run in background with job tracking

Environment Variables

Example with self-signed certificate:

{
  "mcpServers": {
    "design-system": {
      "command": "node",
      "args": ["/path/to/dist/index.js"],
      "env": {
        "STORYBOOK_URL": "https://my-storybook.example.com",
        "NODE_TLS_REJECT_UNAUTHORIZED": "0"
      }
    }
  }
}

Usage

See DEVELOPMENT.md for detailed setup instructions.

Available Tools (9 total)

Core Tools

1. list_components

   • Lists all available components from the Storybook instance
   • Use compact: true for minimal output (reduces response size)
   • Filter by category parameter
   • Supports pagination with page and pageSize (default: 20)

2. get_component_html

   • Extracts HTML from a specific component story
   • Async by default: Returns job_id, use job_status to poll for results
   • Set async: false for synchronous mode (uses timeout parameter)
   • Use variantsOnly: true to get list of available variants (sync, fast)
   • Optional includeStyles: true for CSS extraction (Storybook CSS filtered out)
   • Story ID format: "component-name--story-name" or just "component-name" (auto-resolves to default variant)

3. search_components

   • Search components by name, title, category, or purpose
   • query: Search term (use "*" for all)
   • purpose: Find by function ("form inputs", "navigation", "feedback", "buttons", etc.)
   • searchIn: "name", "title", "category", or "all" (default)
   • Supports pagination with page and pageSize

Component Analysis Tools

4. get_component_dependencies
   • Analyzes rendered HTML to find which other components are used internally
   • Detects React components, web components, and CSS class patterns
   • Requires story ID format: "component-name--story-name"

Design System Tools

5. get_theme_info

   • Extracts design system theme (colors, spacing, typography, breakpoints)
   • Gets CSS custom properties/variables
   • Use includeAll: true for all CSS variables

6. get_external_css

   • DEFAULT: Returns only design tokens + file stats (avoids token limits)
   • Extracts & categorizes tokens: colors, spacing, typography, shadows
   • Use includeFullCSS: true only when you need full CSS content
   • Security-protected: only accepts URLs from same domain as Storybook

Job Management Tools

7. job_status

   • Check status of an async job
   • Returns: status, result (when completed), error (when failed)
   • Poll this after calling get_component_html in async mode

8. job_cancel

   • Cancel a queued or running job
   • Returns whether cancellation was successful

9. job_list

   • List all jobs with their status
   • Filter by status: "all" (default), "active" (queued/running), "completed"
   • Returns job list + queue statistics

Example Usage

// List all components (compact mode recommended)
await list_components({ compact: true });

// Search for components
await search_components({ query: "button", searchIn: "name" });

// Find components by purpose
await search_components({ purpose: "form inputs" });

// Get variants for a component
await get_component_html({
  componentId: "button",
  variantsOnly: true
});
// Returns: { variants: ["primary", "secondary", "disabled"] }

// Get HTML (async mode - default)
await get_component_html({ componentId: "button--primary" });
// Returns: { job_id: "job_xxx", status: "queued" }

// Poll for result
await job_status({ job_id: "job_xxx" });
// Returns: { status: "completed", result: { html: "...", classes: [...] } }

// Get HTML (sync mode)
await get_component_html({
  componentId: "button--primary",
  async: false,
  timeout: 30000
});
// Returns: { html: "...", classes: [...] }

// Get HTML with styles
await get_component_html({
  componentId: "button--primary",
  async: false,
  includeStyles: true
});

// Check all running jobs
await job_list({ status: "active" });

// Extract theme info
await get_theme_info({ includeAll: false });

// Get design tokens from CSS
await get_external_css({
  cssUrl: "https://my-storybook.com/assets/main.css"
});

AI Assistant Usage Tips

1. Start with discovery: Use list_components with compact: true
2. Get variants first: Use get_component_html with variantsOnly: true
3. Use async for HTML: Default async mode prevents timeouts on large components
4. Poll job_status: Check job completion before reading results
5. Search by purpose: Use search_components with purpose parameter

Example Prompts

Once connected, you can use natural language prompts with Claude:

MCP Servers Connected

Component Discovery:

Show me all available button components in the design system

Building New Features:

I need to create a user profile card. Find relevant components
from the design system and show me their HTML structure.

Design System Analysis:

Extract the color palette and typography tokens from the design system.
I want to ensure my new component matches the existing styles.

Component Migration:

Get the HTML and styles for the "alert" component. I need to
recreate it in a different framework while keeping the same look.

Multi-Tool Workflow:

First list all form-related components, then get the HTML for
the input and select components. I'm building a registration form.

How It Works

Connects to Storybook via /index.json and /iframe.html endpoints. Uses Puppeteer with headless Chrome for dynamic JavaScript rendering. Long-running operations use an in-memory job queue with max 2 concurrent jobs and 1-hour TTL for completed jobs.

Troubleshooting

• Ensure Storybook is running and STORYBOOK_URL is correct
• Use list_components first to see available components
• For large components, use async mode (default) and poll job_status
• Check /index.json endpoint directly in browser
• SSL certificate errors: Set NODE_TLS_REJECT_UNAUTHORIZED=0 for self-signed certificates
• See DEVELOPMENT.md for detailed troubleshooting

Requirements

• Node.js 20+
• Chrome/Chromium (for Puppeteer)
• Running Storybook instance (see below for supported versions)

Supported Storybook versions

Storybook 7, 8, 9 and 10. The server reads the story index from /index.json, falling back to /stories.json, and renders stories through /iframe.html?id=<storyId> — endpoints that have been stable across all four major versions.

Storybook 6 and earlier are not supported: they predate /index.json and use a different story-id scheme.

Both a dev server (npm run storybook) and a built static Storybook served over HTTP will work.

Development

See DEVELOPMENT.md for detailed development instructions.

Author

Created by Tomáš Grasl

License

MIT