figma-ui-mcp

Claude Code to Figma · Antigravity to Figma · Cursor to Figma · Any MCP IDE to Figma

_{✅ Tested: Claude Code, Antigravity  |  🔧 Compatible: Cursor, VS Code, Windsurf, Zed (any MCP stdio client)}

Bidirectional Figma MCP bridge — let AI assistants (Claude Code, Cursor, Windsurf, Antigravity, VS Code Copilot, or any MCP-compatible IDE) draw UI directly on Figma canvas and read existing designs back as structured data, screenshots, or code-ready tokens. No Figma API key needed — works entirely over localhost.

Requires Figma Desktop — the plugin communicates with the MCP server over localhost HTTP polling. Figma's web app does not allow localhost network access, so Figma Desktop is required.

Claude ──figma_write──▶ MCP Server ──HTTP (localhost:38451)──▶ Figma Plugin ──▶ Figma Document
Claude ◀─figma_read──── MCP Server ◀──HTTP (localhost:38451)── Figma Plugin ◀── Figma Document

How the localhost bridge works

The MCP server starts a small HTTP server bound to localhost:38451. The Figma plugin (running inside Figma Desktop) uses long polling — the server holds requests up to 25s until work arrives, delivering near-realtime latency. All traffic stays on your machine — nothing is sent to any external server.

Multi-instance support (v2.3.0+): Multiple Figma files/tabs can connect simultaneously. Each plugin instance sends a sessionId, and the bridge routes operations to the correct session. Use the optional sessionId param in figma_write/figma_read to target a specific file.

Features

What's new in v2.4.0

Quick Start

Step 1 — Add the MCP server to your AI client

Choose your platform:

# Project scope (default)
claude mcp add figma-ui-mcp -- npx figma-ui-mcp

# Global scope (all projects)
claude mcp add --scope user figma-ui-mcp -- npx figma-ui-mcp

Edit config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "figma": {
      "command": "npx",
      "args": ["-y", "figma-ui-mcp"]
    }
  }
}

Edit .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "figma": {
      "command": "npx",
      "args": ["-y", "figma-ui-mcp"]
    }
  }
}

Edit .vscode/mcp.json (project) or add to settings.json (global):

{
  "mcp": {
    "servers": {
      "figma": {
        "command": "npx",
        "args": ["-y", "figma-ui-mcp"]
      }
    }
  }
}

Edit ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "figma": {
      "command": "npx",
      "args": ["-y", "figma-ui-mcp"]
    }
  }
}

1. Open "..." dropdown at the top of the agent panel
2. Click "Manage MCP Servers" → "View raw config"
3. Add to mcp_config.json:

{
  "mcpServers": {
    "figma": {
      "command": "npx",
      "args": ["-y", "figma-ui-mcp"]
    }
  }
}

git clone https://github.com/TranHoaiHung/figma-ui-mcp
cd figma-ui-mcp
npm install
# Then point your MCP client to: node /path/to/figma-ui-mcp/server/index.js

⚠️ IMPORTANT: After adding the MCP server, you MUST restart your IDE / AI client (quit and reopen). The MCP server only loads on startup — simply saving the config file is not enough. This applies to Claude Code, Cursor, VS Code, Windsurf, and Antigravity.

Step 2 — Install the Figma plugin

⬇ Download plugin.zip — no git clone needed

1. Download and unzip plugin.zip anywhere on your machine
2. Open Figma Desktop (required — web app cannot access localhost)
3. Go to Plugins → Development → Import plugin from manifest...
4. Select manifest.json from the unzipped folder
5. Run Plugins → Development → Figma UI MCP Bridge

The plugin UI shows a green dot when the MCP server is connected.

Step 3 — Connect AI to Figma

Tell your AI assistant to connect:

"Connect to figma-ui-mcp"

The AI will call figma_status and confirm:

✅ Connected — File: "My Project", Page: "Page 1", Plugin v2.4.0

If you see "Plugin not connected", make sure the Figma plugin is running (Step 2).

Step 4 — Start designing with prompts

Once connected, just describe what you want in natural language:

"Use figma-ui-mcp to draw a login screen for mobile"

The AI will automatically:

1. Call figma_docs to load the API reference and design rules
2. Call figma_read get_page_nodes to understand the current canvas
3. Call figma_write to create the design on your Figma canvas
4. Call figma_read screenshot to verify the result

Prompt examples

Tips for better results

• Be specific about style: "dark theme", "glassmorphism", "minimal white" gives the AI clear direction
• Mention platform: "mobile" (390×844), "tablet" (768×1024), "desktop" (1440×900)
• Iterate: After the first draw, say "fix the spacing" or "make the buttons bigger" — the AI reads and modifies existing nodes
• Use selection: Select a frame in Figma and ask "improve this design" — the AI reads your selection first
• Multi-screen flows: "Now draw the signup screen next to the login screen" — the AI positions frames side by side

Workflow summary

You: "Connect to figma-ui-mcp"
AI:  ✅ Connected to Figma

You: "Draw a mobile onboarding screen with 3 steps"
AI:  [calls figma_docs → figma_write → figma_read screenshot]
AI:  ✅ Done — here's what I created: [inline screenshot]

You: "The title text is not centered"
AI:  [calls figma_read get_selection → figma_write modify → screenshot]
AI:  ✅ Fixed — text is now centered

You: "Now draw the next onboarding screen beside it"
AI:  [reads page_nodes to find position → draws at x+440]
AI:  ✅ Done — 2 screens side by side

figma_status     — check connection (always call first)
figma_docs       — load API reference (call before drawing)
figma_write      — draw / modify UI on canvas
figma_read       — extract design data, screenshots, SVG

Usage Examples

Draw a screen

Ask Claude: "Draw a dark dashboard with a sidebar, header, and 4 KPI cards"

Claude calls figma_write with code like:

await figma.createPage({ name: "Dashboard" });
await figma.setPage({ name: "Dashboard" });

const root = await figma.create({
  type: "FRAME", name: "Dashboard",
  x: 0, y: 0, width: 1440, height: 900,
  fill: "#0f172a",
});

const sidebar = await figma.create({
  type: "FRAME", name: "Sidebar",
  parentId: root.id,
  x: 0, y: 0, width: 240, height: 900,
  fill: "#1e293b", stroke: "#334155", strokeWeight: 1,
});

await figma.create({
  type: "TEXT", name: "App Name",
  parentId: sidebar.id,
  x: 20, y: 24, content: "My App",
  fontSize: 16, fontWeight: "SemiBold", fill: "#f8fafc",
});
// ... continue

Read a design

Ask Claude: "Read my selected frame and convert it to Tailwind CSS"

Claude calls figma_read with operation: "get_selection", receives the full node tree,
then generates corresponding code.

Screenshot a frame

figma_read  →  operation: "screenshot"  →  nodeId: "123:456"

Returns a base64 PNG Claude can analyze and describe.

Architecture

figma-ui-mcp/
├── server/
│   ├── index.js            MCP server (stdio transport)
│   ├── bridge-server.js    HTTP bridge on localhost:38451
│   ├── code-executor.js    VM sandbox — safe JS execution
│   ├── tool-definitions.js MCP tool schemas
│   └── api-docs.js         API reference text
└── plugin/
    ├── manifest.json       Figma plugin manifest
    ├── code.js             Plugin main — operation handlers
    └── ui.html             Plugin UI — HTTP polling + status

Security

Available Write Operations (figma_write)

Core CRUD

Page Management

Node Operations

Components

Design Tokens & Styles

applyVariable supported fields — bind any FLOAT/COLOR/BOOLEAN variable to:
fill, stroke, opacity, width, height, cornerRadius (+ topLeftRadius / topRightRadius / bottomLeftRadius / bottomRightRadius), strokeWeight, paddingTop, paddingBottom, paddingLeft, paddingRight, itemSpacing, counterAxisSpacing, fontSize, letterSpacing, lineHeight, paragraphSpacing, paragraphIndent, visible

Image & Icon Helpers (server-side)

Prototyping & Interactions (v2.4.0)

Supported transitions: SMART_ANIMATE, DISSOLVE, MOVE_IN, MOVE_OUT, PUSH, SLIDE_IN, SLIDE_OUT, INSTANT
Supported easings: LINEAR, EASE_IN, EASE_OUT, EASE_IN_AND_OUT, CUSTOM_BEZIER

Scroll Behavior (v2.4.0)

Variant & Component Swapping (v2.4.0)

Available Read Operations (figma_read)

includeHidden param (boolean, default false) — available on get_selection, get_design, search_nodes, scan_design. When false (default), nodes with visible: false are skipped. Pass true to include hidden layers.

Working with an Existing Project

When opening a Figma file that already has a design system, always read before drawing.

Step 1 — Read what exists

figma_read get_variables      → load variable IDs (Design Tokens)
figma_read get_styles         → load paint/text style IDs and hex values
figma_read get_local_components → load component IDs
figma_read get_page_nodes     → load top-level frame IDs

Step 2 — Build lookup maps in figma_write

// Load variables → build varMap: name → id
var vars = await figma.get_variables();
var varMap = {};
for (var ci = 0; ci < vars.collections.length; ci++) {
  var col = vars.collections[ci];
  for (var vi = 0; vi < col.variables.length; vi++) {
    var v = col.variables[vi];
    varMap[v.name] = v.id;
  }
}

// Load styles → build colorMap: name → hex, textMap: name → {fontSize, fontWeight}
var styles = await figma.get_styles();
var colorMap = {}, textMap = {};
styles.paintStyles.forEach(function(s) { colorMap[s.name] = s.hex; });
styles.textStyles.forEach(function(s)  { textMap[s.name]  = s; });

// Load components → build compMap: name → id
var comps = await figma.get_local_components();
var compMap = {};
comps.components.forEach(function(c) { compMap[c.name] = c.id; });

Step 3 — Create nodes using discovered values

// Use colorMap for fill (hex from existing styles)
var card = await figma.create({
  type: "FRAME", name: "Card",
  fill: colorMap["color/bg-surface"] || "#FFFFFF",
  width: 360, height: 200,
  layoutMode: "VERTICAL", paddingTop: 16, paddingLeft: 16,
  paddingBottom: 16, paddingRight: 16, itemSpacing: 12,
});

// Then bind variables so light/dark mode switches propagate
if (varMap["bg-surface"])
  await figma.applyVariable({ nodeId: card.id, field: "fill",        variableId: varMap["bg-surface"] });
if (varMap["radius-md"])
  await figma.applyVariable({ nodeId: card.id, field: "cornerRadius",variableId: varMap["radius-md"] });
if (varMap["spacing-md"])
  await figma.applyVariable({ nodeId: card.id, field: "paddingTop",  variableId: varMap["spacing-md"] });

Step 4 — Instantiate components with overrides

// Prefer existing components over drawing from scratch
if (compMap["btn/primary"]) {
  await figma.instantiate({
    componentId: compMap["btn/primary"],
    parentId: card.id,
    overrides: { "Label": { text: "Confirm", fill: "#FFFFFF" } }
  });
}

Step 5 — Pin frames to Light / Dark mode

var collection = vars.collections.find(function(c) { return c.name === "Design Tokens"; });

// Duplicate frame and preview both modes side by side
var frames = await figma.get_page_nodes();
var homeId  = frames.find(function(f) { return f.name === "Home"; }).id;
var light   = await figma.clone({ id: homeId, x: 0,    name: "Preview/Light" });
var dark    = await figma.clone({ id: homeId, x: 1540, name: "Preview/Dark"  });
await figma.setFrameVariableMode({ nodeId: light.id, collectionId: collection.id, modeName: "light" });
await figma.setFrameVariableMode({ nodeId: dark.id,  collectionId: collection.id, modeName: "dark"  });

Full API reference and all design rules: run figma_docs in your AI client.

Star History

If figma-ui-mcp helps you, please give it a star — it helps others discover the project!

License

MIT © TranHoaiHung — free to use, modify, and distribute. See LICENSE for details.

Keywords

figma mcp, claude code to figma, cursor to figma, ai to figma, figma ai plugin, figma mcp bridge, figma mcp server, figma design to code, code to figma design, ai ui design, figma automation, figma plugin ai, model context protocol figma, claude figma, windsurf figma, vs code figma, antigravity figma, ai design tool, figma api alternative, figma localhost plugin, draw ui with ai, ai generate figma design, figma design system ai, mcp server figma, figma read design, figma write design, bidirectional figma, figma desktop plugin, npx figma-ui-mcp