@vizejs/musea-mcp-server

MCP (Model Context Protocol) server for Musea design system integration.

Features

• AI Integration - Connect Musea to AI assistants
• Component Discovery - Query available components
• Design Token Access - Retrieve design system tokens
• Documentation - Access component docs via MCP

Installation

Install vp once from the Vite+ install guide, then add the server to your project:

vp install -D @vizejs/musea-mcp-server

Usage

With Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "musea": {
      "command": "vp",
      "args": ["dlx", "@vizejs/musea-mcp-server", "--project", "/path/to/project"]
    }
  }
}

Standalone

vp dlx @vizejs/musea-mcp-server --project ./my-vue-app

MCP Tools

• analyze_component - Extract props and emits from a Vue SFC or linked art component
• get_palette - Derive an interactive props palette from an art file
• list_components - List registered components with status, variants, and resource URIs
• get_component - Get full component details, analysis, palette, and docs
• get_variant - Get one specific variant
• search_components - Search components by title, tags, category, component name, and variants
• recommend_components - Rank components by user intent or UI task
• generate_variants - Generate an .art.vue draft from a component
• generate_csf - Convert an art file to Storybook CSF
• generate_docs - Generate Markdown docs for a component
• generate_catalog - Generate a catalog for the whole design system
• get_tokens - Read design tokens
• search_tokens - Search tokens without loading the full token tree

Reproducible Prompting

Musea helps AI assistants stay grounded in real component data, but reusable prompts still need to be tuned for reproducibility.

A practical loop is:

1. Fix a few evaluation scenarios first.
2. Run the prompt in fresh assistant sessions.
3. Collect both self-reported ambiguity and tool/checklist traces.
4. Fix one ambiguity per iteration.

For Musea-specific workflows, check whether the assistant:

• used real components from the registry
• matched prop and variant names to MCP output
• read tokens before suggesting visual values
• reported missing metadata instead of guessing

The full MCP guide includes a longer reproducibility workflow and prompt template: docs/content/integrations/mcp.md

License

MIT