Figma MCP Go

figma-mcp-go

Figma MCP — Local Plugin Integration

Open-source Figma MCP server with full read and write access through a Figma plugin. Turn text into designs and designs into code. Works with Claude, Cursor, GitHub Copilot, and any MCP client.

Highlights

• Runs locally through the Figma Plugin API. No REST API token needed.
• Reads and edits the open Figma file live — 60 tools.
• Covers styles, variables, components, prototypes, text, and multi-step batches with undo.
• Built-in prompts such as read_design_strategy and design_strategy.
• Lean on tokens: short tool descriptions, capped node trees, and images returned as images.

Styles, Variables, Components, Prototypes, and Content

https://github.com/user-attachments/assets/eae41471-fc72-4574-8261-4f42c38b8c99

Text to Design, Design to Code

https://github.com/user-attachments/assets/17bda971-0e83-4f18-8758-8ac2b8dcba62

Why this exists

Most Figma MCP servers use the cloud Figma REST API. AI tools make hundreds of quick calls per session, and every cloud round trip adds delay.

This server talks to a Figma plugin on your own machine instead. You get fast, live read and write access to the open file, with no API token.

Setup

Install with npx or bunx. No build step. Watch the setup video or follow the steps below.

1. Add the server to your AI tool

Claude Code CLI

# via npx
claude mcp add -s project figma-mcp-go -- npx -y @tunglt1810/figma-mcp-go@latest

# or via bunx
claude mcp add -s project figma-mcp-go -- bunx @tunglt1810/figma-mcp-go@latest

Codex CLI

# via npx
codex mcp add figma-mcp-go -- npx -y @tunglt1810/figma-mcp-go@latest

# or via bunx
codex mcp add figma-mcp-go -- bunx @tunglt1810/figma-mcp-go@latest

.mcp.json (Claude and other MCP-compatible tools)

{
  "mcpServers": {
    "figma-mcp-go": {
      "command": "npx",
      "args": ["-y", "@tunglt1810/figma-mcp-go"]
    }
  }
}

.vscode/mcp.json (Cursor / VS Code / GitHub Copilot)

{
  "servers": {
    "figma-mcp-go": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@tunglt1810/figma-mcp-go"
      ]
    }
  }
}

2. Install the Figma plugin

1. In Figma Desktop: Plugins → Development → Import plugin from manifest
2. Pick manifest.json from plugin.zip
3. Run the plugin in any Figma file

3. Use more than one AI tool (optional)

Each AI tool starts its own server, but only one can hold the plugin connection on a port.

Same file: change nothing. The first server takes port 1994; the others forward their calls to it. If it stops, another takes over in a few seconds.

Different files: give each tool its own port, and set the same port in that file's plugin (settings gear):

{
  "mcpServers": {
    "figma-mcp-go": {
      "command": "npx",
      "args": ["-y", "@tunglt1810/figma-mcp-go", "--port", "1995"]
    }
  }
}

The plugin remembers the last port for all files, so check it when you open the plugin elsewhere.

--ip 0.0.0.0 accepts connections from other machines.

Security: the plugin connection has no login. On the default 127.0.0.1 only your machine can reach it. On any other address, anyone who can reach the port can read and edit your open file. The server warns you and the plugin turns on confirm mode. Prefer an SSH tunnel.

4. Read-only mode (optional)

Add --read-only to offer only the 17 tools that cannot change the Figma file. These are the read tools, export_screenshots, export_frames_to_pdf, and get_image_bytes. The tool list that the AI loads is about 84% smaller.

{
  "mcpServers": {
    "figma-mcp-go": {
      "command": "npx",
      "args": ["-y", "@tunglt1810/figma-mcp-go", "--read-only"]
    }
  }
}

The flag limits one server only. Another AI tool on the same port can still edit the file, unless it also uses the flag.

Plugin panel

The panel shows the connected file, the selection, and what the AI is doing.

Settings and panel size are saved per machine. The panel follows Figma's light and dark theme.

Dev Mode

set_codegen_result saves code on a node. It shows in Dev Mode's Code panel for everyone on the file. Code on a component shows for all its instances. The Language menu filters blocks; if a language has no block, all blocks show.

How it works

The first server to start becomes the leader and holds the only WebSocket to the plugin. Other servers are followers and forward their calls to the leader.

flowchart LR
    client([MCP client]) -- stdio --> tools
    client2([2nd MCP client]) -- stdio --> follower[Follower]

    subgraph leaderProc [Leader process]
        tools[tools<br/>60 tools, Check] -- Send --> leader[cluster.Leader<br/>:1994 /rpc /ws /ping]
        leader --> bridge[bridge<br/>request/response match]
    end

    follower -. POST /rpc .-> leader

    subgraph plugin [Figma plugin]
        ui[Plugin UI<br/>Svelte iframe] -- postMessage --> main[Plugin main<br/>dispatch handlers]
    end

    bridge == WebSocket ==> ui
    main -- read / write --> figma[(Figma document)]

One tool call, end to end:

sequenceDiagram
    participant C as MCP client
    participant T as tools
    participant L as cluster
    participant B as bridge
    participant U as Plugin UI
    participant M as Plugin main
    participant F as Figma

    C->>T: tools/call (stdio)
    T->>L: Send after Check
    Note over L: A follower first sends POST /rpc to the leader
    L->>B: forward
    B->>U: WebSocket request
    U->>M: postMessage
    M->>F: read / write nodes
    F-->>M: nodes
    M-->>U: response
    U-->>B: WebSocket reply
    B-->>L: matched by id
    L-->>T: (any, error)
    T-->>C: tool result

Tools

Read

Export

Create

Edit