Figma Edit MCP

Connect AI assistants to Figma via Model Context Protocol to read designs, create and modify elements, and manage design systems programmatically.

This plugin empowers your AI assistant to become a Figma assistant, executing design updates Safer, Cleaner, and Faster than a human ever could.

This plugin allows you as a Designer to focus purely on creative decision-making, leaving the error-prone & repetitive manual changes to the automated systems.

🏆 Featured by IDEO: Letting AI edit live Figma files without trusting it to behave

📖 Documentation: neozhehan.github.io/figma-edit-mcp

What This Does for You

The project has three goals: Safer, Cleaner, and Faster.

• 🛡️ Safer: The plugin checks every action the AI requests before that action runs. If an action would damage the file, the plugin refuses it and tells the AI why. For example, the plugin will not delete a color variable while layers still use it. Figma itself does not warn you when you delete a variable that is in use — the deletion leaves broken references that are very hard to find afterwards.
• ✨ Cleaner: Because damaging actions are refused, the file does not accumulate broken references or leftover values. Bulk updates cover every layer that matches your request, not just the layers a person remembers to check.
• ⚡ Faster: The AI applies one change across hundreds of layers in seconds. For example, it can replace a product name in every text layer on a page, or copy the overrides from one component instance to fifty others.

These goals support each other: when the plugin refuses mistakes, the file stays consistent, and a consistent file is easier for both you and the AI to work with.

Read the full reasoning behind these goals →

What the AI Cannot Do to Your File

AI assistants sometimes invent things. An AI can name a layer that does not exist, or misremember which layer it was editing. This plugin does not depend on the AI being right. The plugin checks every edit itself, inside Figma, and refuses any edit that fails a check. The AI cannot skip these checks.

What the plugin enforces:

1. The AI can only edit the area you chose. When you connect, you choose one page or one frame for the session. The plugin refuses any edit outside that area. If you choose no area, the AI can read the file but cannot edit any layers.
2. Every edit must name its exact target layer. The plugin compares the layer name the AI supplies with the real layer's name. If the names do not match, the plugin refuses the edit. This stops the AI from editing a layer it invented or misremembered.
3. A bulk edit with one bad item changes nothing. The plugin validates every item in a bulk edit before it changes the first one. If any item fails validation, the plugin rejects the entire bulk edit.
4. Protected layers stay protected. The plugin refuses edits to locked layers and to assets from shared libraries. It also refuses to add, delete, or move layers inside a component instance.
5. The AI cannot delete its own working area. The plugin refuses any action that would delete or replace the page or frame you assigned for the session.
6. Variables and styles require separate permission. Editing your variables or styles requires its own checkbox in the plugin, and both checkboxes are off by default. You grant that permission; the AI cannot grant it to itself.

One honest limit: the plugin checks where an edit happens and which layer it touches. It cannot check whether the edit is the one you wanted. A wrong edit that follows all the rules will go through. So you review the AI's work the same way you would review a colleague's work.

The exact rules, and the conditions under which each one holds, are written down in SAFETY.md. That document is for developers. (Your AI assistant loads the same rules at runtime via the figma-edit skill or the figma-edit://guide/* resources.)

How the Work Is Divided

You do not need to know every Figma feature to direct this system. The work splits three ways:

• You decide what the design should be. You describe the change in plain language, and you judge the result. You do not need to know which Figma menu or panel performs the change.
• The AI assistant turns your request into specific Figma operations and carries out the repetitive work across many layers.
• The plugin runs inside Figma, checks each operation, and refuses the unsafe ones.

The AI contributes scale. The plugin contributes protection. You contribute the design judgment.

Supported AI Integrations

• Cursor
• GitHub Copilot (VS Code)
• Google Antigravity
• Claude Code (VS Code & CLI)
• Claude Desktop (Chat, Cowork & Code)
• LM Studio

Quick Start

The quickest way to run Figma Edit MCP is directly from the NPM registry. You do not need to clone this repository.

1. Configure your AI assistant

Add the server to your AI assistant's MCP configuration:

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

The config file location depends on your host — see Integration-Specific Setup below. Bun users can substitute bunx for npx; both resolve the same package.

2. Start the WebSocket bridge

In a terminal, start the bridge that connects the MCP server to the Figma plugin. Keep this terminal running:

npx -y --package figma-edit-mcp figma-edit-mcp-socket

3. Install the Figma plugin

The Figma plugin ships inside the NPM package. Install the package once to materialize the plugin files on disk:

# In any directory of your choosing (e.g., ~/figma-edit-mcp/)
npm install figma-edit-mcp

Then in the Figma desktop app:

1. Open Plugins → Development → Import plugin from manifest…
2. Select node_modules/figma-edit-mcp/figma_plugin/manifest.json from the directory above.

The plugin is now available under Plugins → Development in any Figma file.

4. (Optional) Install the agent skill

The package also ships a cross-tool skill that teaches your agent the server's safety constraints, error recovery, and tool selection — loaded on demand, so it costs almost nothing until needed. Agents that support the open SKILL.md standard (Claude Code, GitHub Copilot, OpenAI Codex, Cursor, Gemini CLI, Google Antigravity) discover it once it's in their skills directory:

# copy the skill into your agent's skills directory (path varies by host)
cp -R node_modules/figma-edit-mcp/skills/figma-edit ~/.claude/skills/

Clients that don't support skills can still reach the same guidance over the MCP connection as resources under figma-edit://guide/* — no install required.

Running from a local clone? See CONTRIBUTING.md for the contributor-only --local development workflow.

Integration-Specific Setup

Paste the JSON snippet from the Quick Start into your host's MCP config file:

Manual Configuration

If you prefer to edit your host's MCP config directly, paste this snippet into the appropriate config file:

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

Running from a local clone? See CONTRIBUTING.md for the --local workflow.

Contributing

For local development — building from source, running the bridge from a clone, and the --local integrate workflow — see CONTRIBUTING.md.

Windows + WSL Guide

To allow Figma (running on Windows) to connect to the bridge (running inside WSL), the bridge needs to listen on 0.0.0.0 instead of localhost:

npx -y --package figma-edit-mcp figma-edit-mcp-socket --host 0.0.0.0
# or via environment variable:
FIGMA_EDIT_MCP_SOCKET_HOST=0.0.0.0 npx -y --package figma-edit-mcp figma-edit-mcp-socket

Then point the Figma plugin's WebSocket address at your WSL instance's IP.

Usage