Zafiro.FigReader.Mcp

Zafiro.FigReader.Mcp

An offline MCP server that reads local Figma .fig files so an AI agent can inspect and reproduce a design — structure, layout, colors, typography, text and embedded images — without the Figma API, an account or a token. Think Framelink, but it parses the exported .fig binary directly instead of calling Figma's servers.

Why

The Figma REST API is free but requires a token and an internet round-trip, and Framelink-style tools stream large payloads. Zafiro.FigReader.Mcp works on a .fig you exported yourself (Figma → File → Save local copy…), fully offline, and returns compact, token-efficient JSON.

How it works

A .fig file is a ZIP containing canvas.fig (+ meta.json, thumbnail.png, images/, videos/). canvas.fig is Evan Wallace's Kiwi binary format:

"fig-kiwi" (8 bytes) | uint32 version | chunk0 | chunk1 | ...
chunk = uint32 length + payload
chunk0 = Kiwi binary schema (raw DEFLATE) — self-describing
chunk1.. = Kiwi message (raw DEFLATE or Zstandard) → root type "Message"

The message holds a flat list of nodeChanges (Figma's CRDT model); Zafiro.FigReader.Mcp rebuilds the node tree from each node's parentIndex (parent GUID + fractional position). The Kiwi decoder is a clean-room port of evanw/kiwi; Zstandard is handled by ZstdSharp.Port. No native dependencies.

Build & test

dotnet build -c Release
dotnet test                                   # unit tests (no sample needed)
FIGMA_SAMPLE_FIG=/path/to/file.fig dotnet test # also runs the integration tests

Install

Published as a .NET tool on NuGet (requires the .NET 10 SDK/runtime):

# Install globally...
dotnet tool install -g Zafiro.FigReader.Mcp
# ...then run:
figreadermcp

# ...or run without installing:
dnx Zafiro.FigReader.Mcp

The server speaks MCP over stdio (logs go to stderr, so stdout stays clean for the protocol).

MCP client configuration

VS Code (.vscode/mcp.json)

{
  "servers": {
    "figma": {
      "type": "stdio",
      "command": "dnx",
      "args": ["Zafiro.FigReader.Mcp", "--yes"]
    }
  }
}

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "figma": {
      "command": "dnx",
      "args": ["Zafiro.FigReader.Mcp", "--yes"]
    }
  }
}

If you installed the tool globally you can instead use "command": "figreadermcp" with no args.

Build from source

dotnet build -c Release
dotnet src/Zafiro.FigReader.Mcp/bin/Release/net10.0/Zafiro.FigReader.Mcp.dll

(Avoid dotnet run for MCP: its first-run build output can corrupt the stdio protocol.)

Tools

Node ids use the form sessionID:localID. Tools other than load_file default to the most recently loaded file when path is omitted.

Example

// get_node_tree on a button symbol
{
  "id": "5:326", "name": "Button/Primary", "type": "SYMBOL",
  "width": 199, "height": 48, "fills": ["#2730E9"],
  "layout": { "mode": "HORIZONTAL", "gap": 16, "padding": [8,40,8,32],
              "primaryAlign": "CENTER", "counterAlign": "CENTER" },
  "children": [
    { "id": "5:322", "name": "Button Text", "type": "TEXT",
      "text": { "characters": "Button Text", "fontFamily": "SF Pro Display",
                "fontStyle": "Semibold", "fontSize": 18, "color": "#FFFFFF" } }
  ]
}

Layout

src/Zafiro.FigReader.Core      Kiwi decoder, .fig container reader, document model, extraction
src/Zafiro.FigReader.Mcp    MCP stdio server + tools
tests/Zafiro.FigReader.Core.Tests

Limitations

• The Kiwi schema is reverse-engineered; if Figma changes the format, extraction may need updates (the embedded schema keeps basic decoding working across versions).
• Vector path geometry is referenced by blob index and not decoded into SVG (yet).
• Large files take a few seconds to decode on first load (then cached in memory).

License

MIT — see LICENSE.

The Kiwi binary decoder is a port of evanw/kiwi (MIT). Third-party licenses and attributions are listed in THIRD-PARTY-NOTICES.md.

"Figma" is a trademark of Figma, Inc. This project is not affiliated with, endorsed by, or sponsored by Figma, Inc.