4 Ways We’re Using Our MCP Server at Figma | Figma Blog
Justy and Cody dig into Figma's Model‑Context‑Protocol server, unpacking how it lets AI agents edit slides and FigJam boards directly, why custom‑font and asset tools matter, and what real‑world designers should care about.
Transcript
Justy Okay, so I just got back from a coffee run and the Figma MCP server popped up on my radar again—
Cody Right.
Justy —and it feels like the kind of thing that could actually shrink the gap between design mockups and the code that ships, especially now that agents are finally on the canvas.
Cody Yeah.
Justy I spent Thursday wrestling with a stubborn font mismatch in a client deck, so I was pretty relieved to see they finally added custom‑font support to the MCP server. It meant I could just prompt the agent and watch it render the exact typeface I had on my machine, no more fallback to web‑safe approximations.
Cody Mm-hm.
Cody The core claim in the blog is that MCP—Model‑Context‑Protocol—acts as a bridge, pulling design decisions straight into the place where code lives. It serializes the canvas JSON, attaches the current design system token map, and feeds that payload to the language model, so the model can reason about layout constraints without ever seeing the UI. By exposing tools like use_figma and the /figma‑use‑slides skill, the server lets an LLM read the canvas state, generate edits, and p
Justy Exactly.
Justy Mallory’s workflow shows that an evergreen LIVE DECK can be refreshed automatically: the agent pulls the latest release notes from Slack, grabs assets from the design system, then runs the use_figma tool to drop new slides into the template. The first eighty percent of the work is done before a human even opens the file, and the custom fonts keep everything ON‑BRAND.
Cody Sure.
Cody What makes that possible is the new download_assets tool, which streams SVG, PDF, JPG or PNG directly from the file. Because the MCP server now accepts uploaded font files, the rendering pipeline can embed those glyphs at render time, keeping the visual output truly on‑brand.
Justy Speaking of brand, I tried to convince my cat that the AI‑generated slides were a new kind of laser pointer. He stared at the screen for three minutes, then walked away like I’d just broken the internet. I swear he was judging my design choices.
Cody The second workflow pushes that idea into FigJam. By authoring a custom skill and registering it in the community‑resources repo, agents can ingest live spreadsheet data, spin up a wireframe, and lay out sticky notes—all inside the whiteboard. Because FigJam is already collaborative, multiple teammates can watch the agent populate the board and steer it with comments. The skill returns a FigJam node graph that the user can tweak instantly.
Justy Oh interesting.
Justy For designers, that means you can turn a sprint retro into a living board without ever leaving the conversation. The data stays consistent, the visual language stays anchored to the design system, and you avoid the manual copy‑paste nightmare that usually eats up half a day. In practice that shaves a few hours from the sprint cycle, letting the team focus on user testing instead of manual diagramming.
Cody I see.
Cody The only caveat I see is that the approach assumes a mature design system and a stable skill ecosystem. Small teams that still cobble their own components might hit friction when the agent expects a token name that simply doesn’t exist.
Justy So the practical win is real, but you still need a human reviewer to catch edge‑case phrasing or missing assets. It feels less like a magic button and more like a very fast first draft generator.
Cody One thing that impressed me is the batch download mode: you can request a folder of icons and get a zip of SVGs in one call, which cuts down on network chatter when you’re iterating on a component library.
Justy Got it.
Justy If you want to try it yourself, the community‑resources repo on GitHub has a starter skill template, and the Figma docs walk you through registering a custom tool with the MCP server.
Cody I keep hearing people call it the MCP—My Coffee Processor—because it seems to run on caffeine alone.
Justy Alright, I’ll let you get back to your coffee, Cody. Catch you next time when we discover another way to make the design‑to‑code bridge a little less rickety.