The canvas meets the model

The tldraw.dev hero says “Build infinite canvas apps in React with the tldraw SDK,” and lists what is inside: multiplayer with live cursors and viewport following, selection and transformation with nested transforms, a geometry system for hit-testing, theming and screen reader support, a signals library and record store, alignment and distribution, reordering, and camera controls. That is the infrastructure of a drawing tool. What turned it into something else was a weekend in November 2023, which Steve Ruiz tells in make real, the story so far. A day after an announcement, Figma engineer Sawyer Hood posted a video in which he used the tldraw component to draw a user interface, exported it as a picture, sent it to GPT-4V, and got back working HTML and CSS. Two weeks later the open source repository had passed 10,000 GitHub stars.

Make Real is the product of that loop. Draw your interface, select the drawing, click the blue Make Real button, and “in a few seconds, your website will appear on the canvas.” The first design move, Steve writes, was obvious: put the website back on the canvas. Because tldraw’s canvas “has always been made of regular HTML,” it can hold iframes directly, so the generated site is a shape you can move, resize to explore its breakpoints, arrange beside other iterations, and draw on top of. Click Make Real again and the annotations become the next prompt. One problem: an iframe is a black box to the page containing it, so embedded content exports as an empty rectangle and the model cannot see its own previous output. The fix is to hand the previous result’s HTML back with a note to fill in the white box with it. “Like a chat window, the canvas is transformed into a conversation space where you and the AI can workshop an idea together.”

The post is candid about the launch: the rate limit kicked in immediately and broke the app for everyone, so in desperation the team added a text input for users’ own OpenAI keys. It was written “roughly 72 hours” after release, which is to say it is a prototype’s story, told in the same voice as the dragging and rotation posts. It closes with a provocation: “You’ve got a canvas that can hold the whole internet and an AI that can see and think. What will you make?” By 2026 the provocation has a product behind it. Text is not enough opens with “Some ideas just don’t fit in a chat box”: describing an adjustment to part of a design, or a process or workflow, is hard through chat alone. tldraw’s MCP App puts a full canvas inside the conversation; the agent can draw, diagram, and wireframe, and you can edit the canvas too while the agent sees your changes. The canvas has become a surface that people and models share.

Amelia Wattenberger has been arguing for this kind of surface from the other direction. In Evolving the infinite canvas she starts from how linear the web is (top to bottom, or if we are feeling spicy, left to right) and observes that positioning content anywhere “leverages our spatial skills.” Infinite canvases “marry the easy manipulation of the digital world with the spatial strengths of the physical world,” which is why they get used for complex, unsorted, exploratory tasks. But she thinks the concept has been static, and she proposes two kinds of evolution. First, hardcode the behaviours we perform by hand into the canvas itself: preventing overlap, grouping that acts like magnets (“pull fast enough and they detach”). Second, hardcode the Gestalt-like rules we read by eye (similarity, linear order, proximity) into the exported data, so the arrangement you made on the canvas means something to the software. She also questions why an app must choose a text document, a spreadsheet, a timeline, or a canvas at all, since “each interface has its own strengths, weaknesses, and stored information.”

Her essay LLMs as a tool for thought supplies the reason the canvas and the model belong together. Most people, she thinks, use chatbots only as a way to steer somewhere, whereas a tool for thought is “like sketching while thinking”: the point is not creating something but gaining distance from thoughts to evolve them. She describes a cycle of observing, planning, and painting, and a painter who, close to the canvas, loses the view of the whole; a good tool preserves thoughts outside working memory and speeds that cycle up. Chatbots, she says, are like a restaurant with a 20-page menu: good enough for anyone, ideal for nothing. Elsewhere she asks for interfaces that let us “collaborate on tangible artifacts, not just ephemeral chat logs” (Our interfaces have lost their senses). Make Real’s iframes on a canvas are a literal answer to that request. Her broader case against chat as a default is taken up in Data into meaning.

Maggie Appleton supplies the counterweight. Her January 2026 note on Gas Town, Steve Yegge’s entirely vibecoded agent-orchestration system, treats it as speculative design fiction rather than a usable tool, and draws one conclusion that matters here: “When you have a fat stack of agents churning through code tasks, development time is no longer the bottleneck.” Design is: “imagining what you want to create and then figuring out all the gnarly little details required to make your imagination into reality.” Her own holdups are never build time; they are architecture, feel, look, whether a transition is subtle enough, whether the metaphor is right. She adds that language is “a poor medium for designing easing curves and describing aesthetic feelings,” that she always needs to touch the CSS, and that an agent asked to make a diagram struggles because it cannot really see what it is making. Her biggest criticism of Gas Town is that it was not designed ahead of time; as a commenter she quotes put it, “it was vibe designed too.”

Read together, the three voices agree more than they disagree. Steve built a surface where the model’s output is a movable, resizable, drawable object rather than a transcript. Amelia explains why that surface helps thinking. Maggie warns that the thinking is now the scarce part, and that a surface which makes it easy to generate without stopping to look is a footgun. For your own work the lesson is to give models and people the same spatial, visual place to meet, and to design that place so the human keeps the pencil. The canvas is valuable because someone still has to decide what the drawing means.