turbofig

<div align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="docs/wordmark-dark.png"> <source media="(prefers-color-scheme: light)" srcset="docs/wordmark-light.png"> <img alt="turbofig: free Figma bridge for Claude Code, Cursor, Codex and Copilot, no MCP server needed" src="docs/wordmark-light.png" height="80"> </picture> <p><strong>Let AI drive Figma. Instant, lightweight, free.</strong><br> _{No paid Figma seat. No quota. No Figma token. Free and open source, for macOS.}</p>

</div> <!-- DEMO VIDEO: replace with a GitHub user-attachments video URL. Drag the .mp4 into the README in GitHub's web editor to get one. Optional caption below it, only if true: _{Real speed, not sped up.} -->

---

turbofig is a free, open-source Figma bridge for AI agents on macOS. It talks to your agent through plain files, so it needs no MCP server (an MCP mode is there if you prefer). Ask Claude Code, Cursor, Codex or Copilot for real Figma work in plain English, and watch it build in the file you have open.

• Fast, start to finish. Ready the moment your AI starts and already connected, so your first request lands in your file in seconds. Nothing to start, nothing to reconnect, no approval stops in the middle of a job. In a side-by-side design session, it finished in half the time.
• Far lighter on tokens. The same design session used 2.8× fewer tokens. Each screenshot costs about 3× fewer, and your AI learns 4 actions instead of 121. Less of your AI plan goes on overhead, so long sessions go further.
• Free to use. No paid Figma seat, no monthly quota.

_{See how turbofig compares and the benchmark.}

Why turbofig

• Easy. A small native app, with no Node. One Homebrew command installs it, and brew upgrade turbofig updates it. Import the plugin into Figma once, paste the copy-prompt into your agent, and you are done. No Figma token, no sign-in, no shared billing, so a whole team or class can run it side by side.
• Connects by itself, stays connected. Open the plugin and it finds turbofig. No restarting plugins or gateways to get a connection. If the link drops, it reconnects on its own.
• Screenshots stay small. Screenshots are saved as files and shrunk to at most 1200 px, so they do not fill up your AI's memory. Full size is there when your AI asks for it.
• Multi-file. Several files and several agent sessions can run at once, handy when you juggle clients.
• No approval clicks, if your setup asks for them. Some setups ask you to approve every tool call, or block adding MCP servers. turbofig talks to your AI through plain files, so it avoids both and a long job can run without you. See docs/reliability.md.

Why I made it

I build in Figma with AI agents every day, and two things kept getting in the way: speed and cost. The bridges I tried were slow to respond, and they spent a big part of my AI plan just loading their tools before any work began, so long sessions filled up and needed compacting. In my setup they also often needed restarts before they would connect.

turbofig is the bridge I wanted. It is always on and answers straight away, so big jobs move fast. It gives the AI only what it needs, so less of every session goes on overhead. Since switching, I no longer have to compact long design sessions in Claude Code, and I give it big jobs without worrying about the cost.

What you can ask it

• Build a pricing page frame from my existing components.
• Resize one master frame into a full set of ad and social formats.
• Swap copy across dozens of frames and keep every text style intact.
• Rename or restyle hundreds of layers to match a naming rule.
• Create a set of button components with size and state variants.
• Update a text style or a color variable everywhere it is used.
• Make a dark-mode variant of this frame.
• Lay out a slide deck or a storyboard from a text outline.
• Generate a grid of frames from a CSV or JSON list.
• Match a coded component's spacing, colors and type in Figma so design and code agree.

turbofig can run any script a Figma plugin could run, so variables, components, component sets and styles all work. Scriptable too: any process that can write a job file to the bridge folder can drive turbofig, not only a chat agent.

Get started

Takes a few minutes, most of it waiting for Homebrew.

You need an AI agent that can work with files on your Mac, such as Claude Code, Cursor or Copilot CLI.

1. Install. New to the terminal? No coding needed: open the Terminal app, paste these 2 lines, and press Return after each. Need Homebrew first? Get it from brew.sh.

   brew install LukeHawkins/tap/turbofig
   turbofig
   

   This installs and opens turbofig.app and starts it in the background. A tf icon appears in your menu bar. Click it and choose About turbofig… to get started.

   <p align="center"> <img src="docs/menu-bar.png" alt="The turbofig menu-bar menu" width="185" valign="top"> &nbsp;&nbsp; <img src="docs/about-window.png" alt="The turbofig About window" width="442" valign="top"> </p>

2. Add the plugin to Figma (once). In the About window, click Show plugin folder. In Figma Desktop, choose Plugins > Development > Import plugin from manifest…, then drag manifest.json from the Finder window onto the dialog. This menu item exists only in Figma Desktop, not the web app. Then open a file and run turbofig from the Plugins menu once, and leave the panel open.

3. Connect your agent. Click Copy agent prompt in the About window or the plugin panel, and paste it into Claude Code, Cursor, Copilot, or any other agent with local file access. This needs no MCP setup. See docs/agents.md for an agent-neutral version of the same prompt. Chat apps that run in a browser cannot reach your Mac, so they cannot use turbofig.

Prefer MCP (how apps like Claude Code plug in extra tools)? Claude Code: claude mcp add turbofig -- turbofig mcp. See docs/mcp-clients.md for this and other MCP clients.

How turbofig compares

| | turbofig | figma-console-mcp (NPX) | figma-console-mcp (Cloud) | Figma for Agents | |---|---|---|---|---| | Tokens for one design session: a 4-step landing page with screenshots | ✓ 134K | ✗ 375K (2.8× more) | Not tested | Not tested | | Time for that session | ✓ 2.9 min | ✗ 5.9 min, including approval waits | Not tested | Not tested | | Tokens for one screenshot of the same frame | ✓ 1,225 | ✗ 3,264 (2.7× more) | Not tested | Not tested | | Tokens for 4 small jobs, in apps that load every tool (est.) | ✓ ~57K | ✗ ~405K (7× more) | Not tested | Not tested | | Approval prompts in a 4-job test (tester's usual settings) | ✓ 0 | ✗ 6 | Not tested | Not tested | | AI memory the tools take | ✓ 0.3% | ✗ 18% | ✗ 14.5% est. (96 tools) | ✗ 10.3% (41 tools) | | Ready when your AI starts | ✓ Under 0.1 s | ✗ About 2 s (30 s first run) | Hosted, no local start | Hosted, no local start | | Cost and limits | ✓ Free, no limits | ✓ Free | ✓ Free | ✗ Paid Full seat to edit, read quotas, usage-based pricing coming | | Works where MCP servers are blocked, with no MCP approval prompts | ✓ Yes, talks through plain files (an MCP server is there too, if you prefer) | ✗ No, it is an MCP server | ✗ No, it is an MCP server | ✗ No, it is an MCP server | | Needs a Figma access token | ✓ No | ✗ Yes, you create one | ✗ Yes, you create one | Sign in with Figma | | Extra setup | ✓ None after one plugin import | ✗ Node.js, MCP config, Desktop Bridge plugin | ✗ Pair the Desktop Bridge plugin | Connect your AI client | | Runs any Figma plugin script | ✓ Yes | ✓ Yes | ✓ Yes | ~ Partly: no images or custom fonts |

Bottom line: if you work in Figma Desktop, turbofig is the only option here that needs no MCP server, no token and no extra setup. It gives your AI full scripting power with the lightest load, for free.

_{Measured: one run each, same prompts and model; time spent recovering from tool failures is left out on both sides; the cost gap is smaller than the token gap because both re-read a similar fixed context on every call; see Benchmark. est. = the measured work plus each tool list re-read on each of 10 AI calls. Repeated tool lists are cached, so the cost gap is smaller than the token gap, and Claude Code looks tools up on demand once they pass 10% of its memory. AI memory = share of a 200K-token context, counted with an OpenAI tokenizer; Figma for Agents measured through Copilot CLI. Methods and dated sources: docs/comparison.md.}

Benchmark

The same jobs, prompts and model (Claude Sonnet 5), run once with each tool. Every turbofig result was read back from Figma and matched the spec.

<picture> <source media="(prefers-color-scheme: dark)" srcset="docs/benchmark-overview-dark.svg"> <img src="docs/benchmark-overview-light.svg" width="860" alt="Measured side by side. Same prompts, same model (Claude Sonnet 5), one run each, lower is better. Tokens for one design session, a 4-step landing page with a screenshot after each step: turbofig 134,000, figma-console-mcp 375,000 (2.8 times more). Billed cost for that session, in input-token units: turbofig 228,000 total (89,400 for the work, 138,600 for the fixed context), figma-console-mcp 350,000 total, 35 percent more (163,600 for the work, 186,200 for the fixed context). Time for that session: turbofig 2.9 minutes, figma-console-mcp 5.9 minutes including approval waits. Tokens for one screenshot of the same frame at default settings: turbofig 1,225 at 1200 pixels, figma-console-mcp 3,264 at 2000 pixels. Session figures leave out time spent recovering from tool failures, on both sides."> </picture>

• Design session. Build a landing page in 4 steps and check a screenshot after each one. turbofig used 2.8× fewer tokens (134K against 375K) and about a third less in billed cost, and finished in 2.9 minutes against 5.9. Time spent recovering from tool failures is left out on both sides.
• Why. Each figma-console-mcp screenshot cost about 3× more and stayed in the conversation, so every later call re-read it. It also needed more steps: 24 AI calls against 17.
• Small jobs without screenshots. Both cost about the same: the same AI writes the same kind of script. There, turbofig's edge is no approval stops, a start in under 0.1 s, and connecting on the first try.

One run each, so treat the results as a guide. Evidence: design session, recount, screenshot test, results and method, spec check, figma-console-mcp log. The turbofig logs stay private because they contain unrelated personal context.

FAQ

How do I connect Claude Code to Figma?

Install turbofig, import the plugin into Figma Desktop once, th