How Figma Built an MCP Server in Weeks — AI‑Powered Design Meets LLMs

By LatestLLM Editorial ·

How Figma Built an MCP Server in Weeks — AI‑Powered Design Meets LLMs

Figma’s Jesse Lumarie launched an MCP server using React/Tailwind to feed scene‑graph data to LLMs, learning that pixel‑perfect isn’t enough and switching to component pointers for enterprise accessibility.

Ever wonder how a design tool can start talking to AI while you’re still figuring out the spec?

⚡ TL;DR: Figma’s Jesse Lumarie built an MCP server in weeks, using React/Tailwind to feed scene‑graph data to LLMs, learned that pixel‑perfect output isn’t enough and switched to sending component pointers for enterprise‑ready accessibility.

The Simple Version

Think of Figma as a digital canvas that wants to speak the same language as a large language model.

How It Actually Works

Scene Graph Representation

Figma’s canvas is a C++ scene graph that mirrors a DOM.

🎯 Think of it this way: It’s like a LEGO set where each brick knows its shape and color.

Choosing React & Tailwind

They guessed models have seen the most React/Tailwind markup, so they used it to render output.

Why Images Fail

Sending base64 images blew up the context window and got truncated.

From Pixels to Components

Pixel‑perfect buttons aren’t enough; enterprises need accessible, internationalized components.

🎯 Think of it this way: Instead of sending a picture of a button, they send a backstage pass to the real button object.

Automated Evaluation

Hand‑grading evals took hours; now hundreds of automated LLM judges run weekly.

Technical Deep Dive

Three representation paths were evaluated: raw XML, React markup, and a pointer model.

The pointer approach avoids context bloat and preserves design system integrity.

Compatibility matrices were maintained as the MCP spec evolved rapidly.

Why This Matters

Enterprises can now generate UI that respects their own design systems, not just generic pretty outputs.

Faster iteration reduces manual QA and unlocks AI‑driven design workflows.

Key Takeaways

Source: “How Figma Built an MCP Server in Weeks — AI‑Powered Design Meets LLMs” on YouTube by AI Engineer

Tags: AI, LLM, Figma, MCP, Software Development

Related articles