Figma MCP Overview
When to Use
When you want AI agents to generate code directly from Figma designs — using structured layout data, design tokens, and component mappings rather than screenshots alone.
Decision
| If you need... | Use... | Why |
|---|---|---|
| Maximum compatibility, no desktop app | Remote server | Works on all plans, all seats, any editor |
| Selection-based workflow (select in Figma → generate) | Desktop server | Only desktop server can read active Figma selections |
| Lowest friction Claude Code setup | Claude Code plugin | One command install, bundled skills |
| Dev/Full seat on paid plan | Either | Desktop unlocks selection workflow; remote still provides link-based access |
| Starter plan or View/Collab seat | Remote only | Desktop server requires Dev or Full seat on paid plan |
Pattern
The MCP server exposes 13 tools to AI agents. Agents call these tools automatically based on prompts, or you can trigger them explicitly:
# Remote endpoint
https://mcp.figma.com/mcp
# Desktop endpoint (Figma app running)
http://127.0.0.1:3845/mcp
# Verify who is authenticated (remote only)
Prompt: "Who am I in Figma?" → calls whoami tool
Common Mistakes
- Assuming the desktop server works without a paid Dev/Full seat — it requires one
- Using the desktop SSE endpoint (
/sse) instead of the current HTTP endpoint (/mcp) — Figma updated this; use/mcpfor both - Expecting the remote server to read your active Figma selection — it cannot; it needs a link
- Treating MCP output as production-ready code without reviewing against design — treat it as a strong first draft
See Also
- Remote Server Setup | Desktop Server Setup
- Reference: https://developers.figma.com/docs/figma-mcp-server/