# The canvas meets the model · Canvases and tools for thought ### A weekend in November 2023 The [tldraw.dev](https://tldraw.dev) hero says “Build infinite canvas apps in React with the tldraw SDK”: multiplayer with live cursors, selection and transformation with nested transforms, a geometry system for hit-testing, theming and screen reader support, a signals library and record store, alignment, reordering, camera controls. The infrastructure of a drawing tool. > Figure. TSXtldraw.dev quick start [make real, the story so far](https://tldraw.dev/blog/make-real-the-story-so-far) tells what turned it into something else. A day after an announcement, Figma engineer Sawyer Hood posted a video: draw an interface in the tldraw component, export a picture, send it to GPT-4V, get back working HTML and CSS. Two weeks later the repository passed 10,000 GitHub stars. **Make Real** (tldraw: 'Ever wanted to just sketch an interface, press a button, and get a working website? Well, now you can at makereal.tldraw.com.') is that loop as product: draw, select, click the blue button, and “in a few seconds, your website will appear on the canvas.” The first move was to put the website back on the canvas. Because the canvas “has always been made of regular HTML,” it holds iframes, so the result is a shape you move, resize to explore breakpoints, arrange beside other iterations, and draw on; click again and the annotations become the next prompt. An iframe is a black box to its page, so the model cannot see its own output; the fix hands the previous HTML back with a note to fill in the white box. “Like a chat window, the canvas is transformed into a conversation space where you and the AI can workshop an idea together.” ### A surface people and models share The launch was chaos: the rate limit broke the app for everyone, so the team added a text input for users’ own OpenAI keys. Written “roughly 72 hours” after release, the post closes: “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 answer is a product. [Text is not enough](https://tldraw.dev/blog/text-is-not-enough) opens with “Some ideas just don’t fit in a chat box”: an adjustment to part of a design, or a workflow, is hard through chat alone. tldraw’s MCP App puts a canvas in the conversation; the agent draws, diagrams, and wireframes, and you edit too while it sees your changes. Source Steve’s 2026 post [Stay away from my trash!](https://tldraw.dev/blog/stay-away-from-my-trash) is the other side: tldraw began auto-closing external pull requests after an influx of low-quality AI contributions. “In this system, slop is lubrication.” Amelia Wattenberger argues for the surface from the other direction. [Evolving the infinite canvas](https://wattenberger.com/thoughts/evolving-the-infinite-canvas) starts from how linear the web is (top to bottom, or if we are feeling spicy, left to right): infinite canvases “marry the easy manipulation of the digital world with the spatial strengths of the physical world,” hence their use for complex, unsorted, exploratory tasks. She proposes hardcoding the behaviours we do by hand into the canvas (preventing overlap, magnet-like grouping where “pull fast enough and they detach”) and the Gestalt-like rules we read by eye (similarity, linear order, proximity) into the exported data, and asks why an app must choose document, spreadsheet, timeline, or canvas when “each interface has its own strengths, weaknesses, and stored information.” [LLMs as a tool for thought](https://wattenberger.com/thoughts/llms-as-a-tool-for-thought) gives the reason canvas and model belong together. Most people use chatbots only to steer somewhere; a tool for thought is “like sketching while thinking,” gaining distance from thoughts to evolve them. A painter close to the canvas loses the whole; a good tool keeps thoughts outside working memory and speeds the cycle of observing, planning, painting. Chatbots are a restaurant with a 20-page menu, good enough for anyone, ideal for nothing. In [Our interfaces have lost their senses](https://wattenberger.com/thoughts/our-interfaces-have-lost-their-senses) she asks to “collaborate on tangible artifacts, not just ephemeral chat logs”; Make Real’s iframes are a literal answer. Her case against chat as default continues in [Data into meaning](https://learn-ui.com/chapters/explaining/data-into-meaning). ### Design is the bottleneck Maggie Appleton is the counterweight. Her January 2026 note on [Gas Town](https://maggieappleton.com/gastown), Steve Yegge’s entirely vibecoded agent-orchestration system, treats it as speculative design fiction, not a usable tool, and concludes: “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 architecture, feel, look, whether a transition is subtle enough. Language is “a poor medium for designing easing curves and describing aesthetic feelings”; she always touches the CSS, and an agent asked for a diagram cannot really see what it makes. Her biggest criticism is that Gas Town was not designed ahead of time; as a commenter she quotes put it, “it was vibe designed too.” The three voices agree. Steve built a surface where the model’s output is a movable, drawable object, not a transcript. Amelia explains why that helps thinking. Maggie warns that thinking is now the scarce part. Give models and people the same spatial place to meet, and design it so the human keeps the pencil. Someone still has to decide what the drawing means.