figma-mcp-poor

A Figma MCP (Model Context Protocol) bridge that lets AI assistants read design context from Figma. Uses the Figma Plugin API (local, no network) instead of the REST API — no API keys, no rate limits, no blown context windows.

AI IDE  ←—MCP stdio—→  MCP Server (Node.js)  ←—WebSocket—→  Figma Plugin (figma.* API)

Quick Start

1. Set up your MCP client

Add the server to your MCP client config (see Client Configuration below).

2. Install the Figma plugin

Option A — Download pre-built (recommended):

1. Go to Releases and download figma-plugin-vX.Y.Z.zip
2. Unzip somewhere stable on your disk
3. In Figma: open any design file → Plugins → Development → Import plugin from manifest... → select manifest.json from the unzipped folder

Option B — Build from source:

git clone https://github.com/whoistus/figma-mcp-poor
cd figma-mcp-poor
npm install
npm run build:plugin

Then in Figma: Plugins → Development → Import plugin from manifest... → select figma-plugin/manifest.json.

Run the plugin — it will connect to the MCP server via WebSocket on localhost:3055.

3. Start designing with AI

Select elements in Figma and use your AI assistant to inspect them. The recommended first call is get_dev_summary — it returns structure, text content, colors, components, and a screenshot in one request.

MCP Client Configuration

Claude Code

In your project's .mcp.json or global ~/.claude.json:

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

Cursor

In .cursor/mcp.json:

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

VS Code (Copilot)

In .vscode/mcp.json:

{
  "servers": {
    "figma": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "figma-mcp-poor"]
    }
  }
}

Windsurf

In ~/.codeium/windsurf/mcp_config.json:

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

Custom port

Set the FIGMA_MCP_PORT environment variable to change the WebSocket port (default: 3055):

{
  "mcpServers": {
    "figma": {
      "command": "npx",
      "args": ["-y", "figma-mcp-poor"],
      "env": {
        "FIGMA_MCP_PORT": "4000"
      }
    }
  }
}

Note: If you change the port, you also need to update the Figma plugin's manifest.json → networkAccess.devAllowedDomains to match.

Available Tools

Tool	Description	Token Cost
get_dev_summary	Start here. Complete summary: structure, texts, colors, components, screenshot	Medium
get_document_info	File name, pages, current page	Low
get_selection	Selected nodes with CSS properties, optional text flattening and screenshot	Low–Medium
get_node_by_id	Inspect a specific node by ID with CSS properties. Paginates large children lists	Low–Medium
get_styles	Local paint/text/effect/grid styles as CSS values	Low
get_variables	Design tokens/variables (colors, numbers, strings, booleans)	Low
get_components	Local components with property definitions	Low–Medium
get_design_context	Design brief (markdown) or detailed context (JSON) for a node	Medium
get_flows	Prototype flows: starting points, interactions, navigation connections	Low–Medium
get_screenshot	Export a node as PNG, JPG, or SVG	Medium
get_user_flow_context	Structured FlowGraph JSON: nodes, edges, BFS paths, explicit no-flow signal	Low–Medium

How It Works

1. The MCP server starts a WebSocket server on localhost:3055
2. The Figma plugin connects to this WebSocket from inside Figma
3. When an AI tool is called, the server sends a request through WebSocket to the plugin
4. The plugin reads data using the Figma Plugin API (figma.*) and sends it back
5. The server formats raw Figma data into developer-friendly CSS-like output and returns it to the AI

Key design decisions:

• Default depth=1 for node traversal prevents context explosion. Use get_node_by_id to drill deeper.
• Formatting is server-side — the plugin sends raw data, the server transforms it. Formatting changes don't require reloading the plugin.
• Smart truncation — responses exceeding the token budget are truncated with hints guiding the AI to drill deeper.
• Multi-instance safe — the first MCP process to start becomes the primary (owns port 3055 + proxy port 3056). Subsequent processes (e.g. Codex starting alongside Claude Code) automatically become secondaries and forward requests through the primary. Only one Figma plugin connection is ever active.

Development

# Install dependencies
npm install

# Build everything
npm run build

# Dev mode (MCP server with auto-reload)
npm run dev:server

# Watch mode (Figma plugin)
npm run watch -w figma-plugin

Troubleshooting

Plugin won't connect

• Make sure the MCP server is running before starting the plugin
• Check that port 3055 (or your custom port) isn't in use: lsof -i :3055
• The plugin only connects to localhost — no remote connections

"No selection" responses

• Select at least one element in Figma before calling selection-based tools
• Make sure the plugin UI is open (the WebSocket client runs in the plugin UI)

WebSocket errors

• Only one Figma file can connect at a time (single-connection design)
• If you switch files, restart the plugin in the new file

Multiple AI tools conflict / port already in use

• Starting in v0.4.0, multiple MCP instances co-exist automatically. The first instance owns port 3055; subsequent ones proxy through it on port 3056.
• If you see EADDRINUSE errors, an old server process may still be running: lsof -i :3055 -i :3056 to find and kill it.

Large designs timing out

• Use get_node_by_id with specific node IDs instead of scanning entire pages
• Reduce depth parameter (default is 1, max is 5)
• Screenshots are capped at 800px width by default — use max_width to adjust

Changelog

0.4.0 (2026-05-05)

• New tool: get_user_flow_context — structured FlowGraph JSON for deterministic AI navigation analysis
  • Returns nodes, edges, paths (BFS from prototype starting points), and metadata.hasFlow
  • Explicit no-flow signal (hasFlow: false, empty edges/paths) when no prototype connections exist — no hallucination
  • Optional includeInference: true adds heuristic-inferred edges (action-keyword buttons, list→detail patterns) when no real flow exists, clearly separated from real edges
  • maxDepth param (default 10, max 20) controls BFS traversal depth; paths capped at 200 to protect token budget
• Multi-instance support — run Claude Code + Codex (or any two MCP clients) simultaneously without port conflicts
  • First process becomes primary: owns Figma plugin WS on port 3055 + proxy listener on port 3056
  • Subsequent processes become secondaries: connect to proxy port, forward all tool calls through the primary
  • No changes to existing tool behaviour; no manual configuration required

0.3.0 (2026-04-15)

Fixes for fetching large design areas with minimal data loss.

• Pagination on get_node_by_id — new childrenOffset and childrenLimit params let AI fetch children in chunks for large frames. Response includes pagination: { offset, limit, total, hasMore }.
• Raised children cap — MAX_CHILDREN bumped from 100 → 200 per node. Fewer silent truncations on real-world designs.
• Fixed componentName bug — get_dev_summary previously reported the instance's node name as the component name. Now correctly resolves the master component name via getMainComponentAsync(). Renamed instances now group properly: PrimaryButton x5: "Save", "Cancel", ....
• Colors inside instances — get_dev_summary.colors now walks into instances for fill colors (previously missed custom fill overrides).
• Configurable depth on get_dev_summary — was hardcoded to 2, now accepts depth param (default 2, max 5) for deeply nested screens.
• Cap on flatten_text — 500-entry default cap prevents token budget blow-ups on text-heavy pages.
• Cap on get_flows — new limit param (default 200, max 1000) with clear truncation hint in output.
• get_design_context brief truncation signal — 16000-char cap now includes an explicit hint instead of silently cutting mid-content.

0.2.0 (2026-04-15)

• New tool: get_flows — prototype flow support
  • Returns flow starting points (page.flowStartingPoints)
  • Collects all prototype connections from node trees
  • Supports all trigger types (click, hover, press, drag, mouse enter/leave, timeout)
  • Supports all action types (navigate, swap overlay, open URL, back, close, set variable)
  • Includes transition animations (dissolve, smart animate, slide, push, etc.) with duration and easing
• Node-level reactions — every serialized node now includes reactions[] when it has prototype interactions, visible in get_node_by_id, get_selection, and other node tools
• Compact arrow format for interactions: Button --click--> Login Screen (smart-animate 300ms)

0.1.0 (2026-04-15)

• Initial public release
• 9 MCP tools: get_dev_summary, get_document_info, get_selection, get_node_by_id, get_styles, get_variables, get_components, get_design_context, get_screenshot
• CSS-like formatting for colors, typography, layout (flexbox), spacing
• Smart truncation with drill-deeper hints
• Depth-controlled serialization with auto depth bonus for COMPONENT_SET (+2) and INSTANCE (+1)
• Screenshot export with configurable max width
• Text flattening and color collection utilities

License

MIT