A local MCP server that lets AI coding agents (Claude, Codex, GitHub Copilot) read and edit the file you have open in Figma Desktop. It talks to Figma through a small development plugin instead of the REST API or Figma's hosted MCP, so:
- it works on a View / Starter seat (the hosted MCP gives a handful of calls a month there),
- there are no rate limits and no response truncation,
- the agent can write to the file, not only read it (via
figma_run_script).
AI agent ──stdio──> figma-bridge (server.js) ──ws://127.0.0.1:3055──> plugin/ui.html ──postMessage──> plugin/code.js (Plugin API)
The agent starts the server (npx -y @maxsbond/figma-bridge) as a stdio MCP server. The server opens a WebSocket on
127.0.0.1:3055 and waits. The "Figma Bridge" plugin, running in Figma Desktop,
connects to it and executes each request with the Plugin API in the open file.
- Node.js 18 or newer (the LTS installer is fine)
- Figma Desktop (the browser version can't reach localhost from a dev plugin)
- Edit access to the file. Figma doesn't run plugins in view-only files; if you only have view access, duplicate the file to your drafts.
Paste this into Claude Code, Codex, Copilot or any other agent that can run terminal commands. It sets up everything except the two clicks in Figma, and tells you what those are.
Install Figma Bridge (https://github.com/MaxsBond/figma-bridge) for me:
1. Add the MCP server to your own MCP config: command `npx`, args `["-y", "@maxsbond/figma-bridge"]`. If the config file already exists, merge the entry instead of overwriting it.
2. Run `npx -y @maxsbond/figma-bridge setup` to copy the Figma plugin to ~/Documents/Figma Bridge Plugin.
3. Tell me how to import the plugin in Figma Desktop (Plugins → Development → Import plugin from manifest…, pick manifest.json from that folder) and how to run it in my file (Plugins → Development → Figma Bridge).
4. Tell me if I need to restart you. After the restart, call figma_status to check the connection.
Prefer to do it by hand? Follow the steps below or the step-by-step guide.
There are two parts: the MCP server, which your AI agent starts on its own, and the
Figma plugin, which you import into Figma once. The server is published on npm as
@maxsbond/figma-bridge, so
there is nothing to clone or build.
New to this? The step-by-step guide walks through the same setup with screenshots and tells you what to say to your agent.
VS Code (GitHub Copilot): click Install in VS Code and confirm in VS Code. Other agents are in the sections below.
You import the plugin once. After that you only need to run it (step 4).
1. Get the plugin files. Run this in a terminal:
npx -y @maxsbond/figma-bridge setupIt copies the plugin to Documents/Figma Bridge Plugin and opens that folder. Run it
again after updating to get the new plugin version.
2. Import the plugin. In Figma Desktop open any design file, click the Figma logo in the top-left corner and go to Plugins → Development → Import plugin from manifest… (right-clicking the canvas gets you the same Plugins menu).
3. Pick manifest.json in the Figma Bridge Plugin folder and click Open.
"Figma Bridge" now shows up under Plugins → Development.
4. Run the plugin whenever you want the agent to work in Figma: open the file and pick Plugins → Development → Figma Bridge.
5. Check the connection. The small plugin window says connected to MCP server once an agent with this MCP is running (see the next sections). If the agent isn't running yet, it says "disconnected — retrying…" and connects by itself when the agent starts. Keep the window open while you work, since closing it stops the bridge.
If you'd rather run a checkout: clone the repo, run npm install, import
plugin/manifest.json from it, and in the configs below use node with
/absolute/path/to/figma-bridge/server.js instead of npx -y @maxsbond/figma-bridge.
claude mcp add --scope user figma-bridge -- npx -y @maxsbond/figma-bridgeCheck it with claude mcp list, or /mcp inside a session. Exports go to
./figma-exports/ in the directory you started claude from.
Or install it as a plugin, which adds the MCP server and the skill in one go. Inside a Claude Code session:
/plugin marketplace add MaxsBond/figma-bridge
/plugin install figma-bridge@figma-bridge
Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or
%APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"figma-bridge": {
"command": "npx",
"args": ["-y", "@maxsbond/figma-bridge"],
"env": {
"FIGMA_BRIDGE_OUT_DIR": "/absolute/path/to/figma-exports"
}
}
}
}Restart Claude Desktop. Set FIGMA_BRIDGE_OUT_DIR here because a desktop app doesn't
start the server in a folder you'd want exports in.
Works for the Codex CLI, the Codex IDE extension and the ChatGPT desktop app. They all
read ~/.codex/config.toml.
codex mcp add figma-bridge -- npx -y @maxsbond/figma-bridgeor add it by hand:
[mcp_servers.figma-bridge]
command = "npx"
args = ["-y", "@maxsbond/figma-bridge"]
# Codex's default per-tool timeout is 60 s; big exports and scripts can take longer.
tool_timeout_sec = 180
# optional
# [mcp_servers.figma-bridge.env]
# FIGMA_BRIDGE_OUT_DIR = "/absolute/path/to/figma-exports"Check it with codex mcp list, or /mcp inside the Codex TUI.
The Install in VS Code link above adds the server to your user settings. To add it
to one project instead, create .vscode/mcp.json:
{
"servers": {
"figma-bridge": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@maxsbond/figma-bridge"],
"env": {
"FIGMA_BRIDGE_OUT_DIR": "${workspaceFolder}/figma-exports"
}
}
}
}Click Start above the server entry (or run MCP: List Servers → figma-bridge →
Start), switch Copilot Chat to Agent mode and make sure the figma_* tools are
ticked in the tools picker. MCP tools aren't available in Ask/Edit mode. On Copilot
Business/Enterprise your organization admin has to allow MCP servers.
Add it with /mcp add inside copilot, or edit ~/.copilot/mcp-config.json:
{
"mcpServers": {
"figma-bridge": {
"type": "local",
"command": "npx",
"args": ["-y", "@maxsbond/figma-bridge"],
"tools": ["*"]
}
}
}| Tool | What it does |
|---|---|
figma_status |
Connection check: file name/key, current page, selection |
figma_list_pages |
Pages of the open file |
figma_get_tree |
id / name / type / box outline to a given depth, for finding node ids |
figma_export_json |
Node JSON to disk: rest (same shape as the REST API, JSON_REST_V1), compact (trimmed, with style/variable/component names resolved) or auto |
figma_export_image |
PNG / JPG / SVG / PDF render to disk |
figma_export_image_fills |
Original raster images used in image fills |
figma_run_script |
Arbitrary Plugin API JavaScript (read and write), figma in scope, top-level await, use return |
Node ids can be given as 1:2, 1-2 or a full Figma URL with ?node-id=.
Results over ~15 KB are saved to a file and the tool returns the path instead.
Things to try once it's connected:
- "What's selected in Figma? Export it as PNG and JSON."
- "Build this React component from Figma node 12:345."
- "Rename every layer called
Frame 12…on this page to something meaningful." - "Make a dark variant of the selected component set."
skills/figma-bridge/ is a skill that tells the agent
when to reach for the figma_* tools and how to use them well: check the connection
first, find nodes cheaply, export instead of dumping JSON inline, write safe Plugin API
scripts, and build code from a frame. It also walks the agent through setup when the
server or plugin is missing.
- Claude Code: the plugin install above includes it. Or copy the folder to
~/.claude/skills/figma-bridge/. - Other agents that support skills (Codex, Cursor, Copilot and others): copy the folder to wherever your agent loads skills from.
- Zip for skill marketplaces:
npm run pack:skillbuildsdist/figma-bridge-skill.zipwithSKILL.mdat the root.
| Env var | Default | Meaning |
|---|---|---|
FIGMA_BRIDGE_OUT_DIR |
<cwd>/figma-exports |
Where exports are written (<dir>/<file name>/…) |
FIGMA_BRIDGE_PORT |
3055 |
WebSocket port. If you change it, change it in plugin/manifest.json (devAllowedDomains) and plugin/ui.html (PORT) too |
FIGMA_BRIDGE_TIMEOUT_MS |
120000 |
How long to wait for the plugin to answer one request |
"Figma plugin is not connected": run Plugins → Development → Figma Bridge in the file you want to work on. The plugin only sees that one file.
"Port 3055 is already taken": only one server can hold the port, and the plugin talks to whoever does. This happens when two agents run at once (say Claude Code and Codex both have the MCP enabled) or an old server process is still alive. Find it with
lsof -iTCP:3055 -sTCP:LISTENand close that agent or kill the process. The next figma_* call retries the bind, and
the plugin reconnects within ~10 s. To run two agents at the same time, give one of them a
different FIGMA_BRIDGE_PORT and a second copy of the plugin pointing at that port.
Scripts fail with "Cannot unwrap symbol": the result contains figma.mixed
(for example the fontSize of a text node with mixed sizes). Convert it before returning,
e.g. JSON.parse(JSON.stringify(value, (k, v) => typeof v === 'symbol' ? 'MIXED' : v)).
Node not found / page not loaded in scripts: the plugin uses
documentAccess: dynamic-page, so use the async APIs (figma.getNodeByIdAsync,
node.getMainComponentAsync(), page.loadAsync() / figma.loadAllPagesAsync()).
figma.currentPage is whatever page is open in Figma right now, so address nodes by id
rather than relying on it.
The WebSocket listens on 127.0.0.1 only and accepts connections only with
Origin: null (the plugin's sandboxed iframe), so regular web pages can't talk to it.
Keep in mind that figma_run_script runs whatever code the agent sends with full edit
rights on the open file. Figma's undo works for it, but review what your agent is doing
in files that matter.




