FigOCR
March 19, 2026
FigOCR started life as "Screenshot to Layout" — a Figma plugin that promised to turn screenshots into layouts. I kept the plugin free for a long time as part of a deal in the Figma Creator program, and later because I felt that, while it delivered some value, the plugin did not live up to its name.
The problem was that the plugin did not up to its name fully. This frustrated me a lot.
I wanted to build this kind of conversion plugin that would take screenshots, and turn them into Figma designs.
I had spent a lot of time trying to get it to work (and learning a ton in the process!)
When I did user tests, people asked me if the result could be auto layouted, or use their design system.
I studied OCR methods and I dove into some Python scripts to recognize patterns in images, but did not get very far. What was not very motivating was that from my research, top scientists didn't have a real solution to some of the problems either.
Most logic was based on edge detection. I was ready to give up on Screenshot to Layout, somewhere in November last year.
At an event I explained a person the problem and some underlying stats: 20,000 duplicates, 6,000 accounts created, but only 100-ish active users. More of a niche plugin than anything else.
He asked me about my running costs and I said that it was just Supabase ($25/mo) and around €5-10 in OCR costs per month from Azure.
He encouraged me to at least use it as a lead magnet for whatever I was selling at Obra Studio. The number aren't big, but it's traffic nonetheless, and clearly some people are getting value out of the plugin.
Lately, with the new AI models like Opus 4.6, I've been feeling a bit braver when it comes to development than I used to be.
Over the past few "hack" nights I've been experimenting with building an MCP for Screenshot to Layout.
It started when I heard about this Figsor project, and tried it out. With the code already being on my drive, it was one prompt away to start making it my own.
I tried backporting Obra Autodocs into it - but as an MCP command - and lo and behold, it worked.
I got a bit more ambitious and started asking it to integrate parts from Screenshot to Layout.
I had always held back on integrating an AI layer into Screenshot to Layout since I felt it would be too unreliable, but with the new models being so good at coding I decided - why not.
And again… lo an behold… it works — it can convert screenshots into actual Figma layouts, not just text extraction.
I ran into a problem though. I burned through $100 in a single day using Claude to power the agent. I used an Anthropic API key.
I was thrilled that the results were real: you feed it a screenshot, and you get back a proper Figma layout with positioned elements, not just a pile of text nodes.
After learning that you have much more tokens on an plan (as opposed to the API) I decided to make the MCP logic work directly with people's Claude Pro/Max plans.
It seems to be popular among my AI-loving peers to have such a subscription. Since Figsor worked this way to begin with, this was rather easy to get working.
Another night I went on a whole run trying to improve the OCR layer (seems hard to find something better than Azure); trying if OpenAI models would work better than Claude (they worked more or less the same); with little to no results.
The irony isn't lost on me that I renamed the plugin from "Screenshot to Layout" to "FigOCR" because the layout part wasn't working well enough.
Now that I've got layouts actually working via an AI agent, the old name would have been more accurate than ever.
But, this version has not been released yet.
Right now I'm improving the tools so the agentic AI is better at recreating layouts. The big item on the roadmap: making it design-system aware. That's a huge challenge, but with some human help in the loop, I believe it's possible.
Imagine pointing Screnshot to Layout at a screenshot and getting back a layout that uses your actual design system components — your buttons, your cards, your spacing tokens. That's where we're headed.
While we're building these capabilities into the plugin, we're already offering this as a service. If you have a large number of screenshots that need to be converted into editable Figma layouts, we can do that for you today using our more advanced tooling — a combination of OCR and agentic AI with a human always in the loop for quality control.
Somewhere out there is a design team that needs 20-30 screens converted to editable layers with a 1-day turnaround. We can do that.