Skip to main content
MCP Apps are rich, interactive UI panels that render directly inside your AI assistant. Instead of receiving plain text responses, you get fully interactive applications — with search, filtering, click-to-copy, and real-time data refresh.
Experimental — MCP Apps use the ext-apps protocol extension and require an MCP client with ext-apps support. The protocol may evolve.

How MCP Apps Work

MCP Apps extend the Model Context Protocol with a UI layer. When an AI tool returns a result, it can also include an interactive HTML application that renders inline in the conversation. Key concepts:
  • Server-side tools register with the MCP server and handle data fetching
  • App-only tools are hidden from the AI model and callable only from the UI
  • Resource URIs point the host to the HTML application to render
  • App SDK (@modelcontextprotocol/ext-apps) handles the lifecycle, theming, and communication

Available Apps

Token Browser

Browse all design tokens from your Figma file in an interactive table. Features:
  • Organized by variable collection with expandable sections
  • Filter by type (Colors, Numbers, Strings) and search by name
  • Per-collection mode columns (Light, Dark, Custom) matching Figma’s layout
  • Color swatches with hex/rgba values and alias reference resolution
  • Click any token name or value to copy it to your clipboard
Usage:
The AI calls figma_browse_tokens which returns a brief summary to the conversation and renders the full interactive Token Browser inline. The tool auto-detects the current file from your connected browser tab — no URL required.

Design System Dashboard

Audit your design system with a Lighthouse-style health scorecard rendered inline. Features:
  • Overall weighted score (0–100) with six category gauges: Naming, Tokens, Components, Accessibility, Consistency, Coverage
  • Expandable category sections with individual findings, severity indicators, and actionable details
  • Diagnostic locations linking findings to specific variables, components, or collections
  • Tooltips explaining each check’s purpose and scoring criteria
  • Refresh button to re-run the audit without consuming AI context
Usage:
The AI calls figma_audit_design_system which returns a brief text summary and renders the full interactive dashboard inline. The tool auto-detects the current file from your connected browser tab — no URL required. The dashboard scores six categories:

Enabling MCP Apps

MCP Apps are gated behind an environment variable and only available in Local Mode.
When ENABLE_MCP_APPS is not set or set to "false", no app tools are registered and existing functionality is completely unaffected.

Architecture

Each MCP App follows a modular structure:
Build pipeline:
  1. TypeScript compiles server.ts to dist/apps/<app-name>/server.js
  2. Vite bundles the UI into a single self-contained dist/apps/<app-name>/mcp-app.html
  3. The server registers a resource URI that serves the built HTML file
  4. The host renders the HTML inline when the tool is called
Apps use the official @modelcontextprotocol/ext-apps SDK for lifecycle management, host theming, and server communication.

Building Apps

The build system supports multiple apps via the APP_NAME environment variable:
Adding a new app requires:
  1. Create src/apps/<name>/server.ts with tool and resource registration
  2. Create src/apps/<name>/ui/mcp-app.html and mcp-app.ts
  3. Add a cross-env APP_NAME=<name> vite build entry to the build:apps script
  4. Register the app in src/local.ts inside the ENABLE_MCP_APPS gate

Roadmap

Planned MCP Apps:
  • Component Gallery — Visual browser for searching and previewing components with variant exploration
  • Style Inspector — Interactive panel for exploring color, text, and effect styles with live previews
  • Variable Diff Viewer — Side-by-side comparison of token values across modes and branches

Learn More