@neur0tx/fig-mcp-server

Local MCP server that reads Figma .fig archives from mockups/ and exposes them as MCP resources plus extraction tools.

Features

• Stdio MCP server (node index.js)
• Auto-discovers every *.fig file in mockups/
• Exposes each file as a resource URI: fig://<fileName>.fig
• Resource reads return full structured JSON context:
  • metadata, node tree, components/component sets, styles, variables, image manifest, indexes

Usage

Put your *.fig files into a local directory (for example mockups/) and configure Codex to run this MCP.

Codex Configuration

Add this MCP server to ~/.codex/config.toml (replace /path/to/mockups with your directory):

[mcp_servers.fig_mcp]
command = "npx"
args = [
  "-y",
  "@neur0tx/fig-mcp-server@latest",
  "--mockups-dir",
  "/path/to/mockups"
]

Mockups directory

Can be anything that's accessible by MCP server (it's probably running inside sandbox). Put all .fig files under that directory:

mockups/

Example:

mockups/mockup.fig
mockups/settings-screen.fig

Server should be pointed to the mockups directory just like in the Codex config above. The parameter itself works like this:

• --mockups-dir <path>: explicit directory with .fig files.
  • If set and invalid/unreadable, server exits immediately with an error.
  • If set and valid/readable, server uses it even if empty.
  • If not set, server resolves directory automatically:
    1. <script_dir>/mockups is used only when it is non-empty (ignoring .gitkeep).
    2. Otherwise it tries <cwd>/mockups.

It's not necessary to restart the server after adding a new file.

Notes and limitations

This is essentially a glorified .fig-file parser and it's mostly vibe-coded too. Information about some of the mockup contents is best-effort and may be partially inaccurate or incomplete. Notable examples of that:

• download_figma_images and get_screenshot prefer archive imageRef bytes when available. If they're missing, the server performs a best-effort node render fallback via third-party renderer which may produce borked screenshots or use incorrect font typefaces.
• For opaque binary payloads, this implementation returns metadata and warnings unless payload is JSON-compatible.
• Component libraries / UI Kits won't be available, only the data from source files will be.

In other words, your mileage may vary. My experience shows that this MCP provides enough data for the agent to construct something that looks like the mockup itself but not quite the same. Don't hesitate to just screenshot things from Figma itself because official Figma MCP does that too (they also have get_screenshot tool).

Other documents

• MCP API description