tldraw

tldraw is a London-based startup founded by Steve Ruiz that builds an infinite-canvas SDK for web applications. Its AI work ranges from Make Real, which turns sketches into working website prototypes, to an interactive canvas for Cursor agents and a desktop application that agents can edit and script locally. The foundation is a canvas built with ordinary web content and React components, allowing interactive websites and widgets to sit alongside drawings, text, and annotations.[7][2][8][10][1]

Make Real began with an outside developer’s experiment. After OpenAI released the GPT-4V API at Dev Day in November 2023, Figma engineer Sawyer Hood demonstrated drawing an interface with tldraw, exporting it as an image, and sending it to GPT-4V to obtain HTML and CSS for a working website. The tldraw team forked his repository the next morning. Its first change was to place the generated website directly on the canvas, using an iframe supported by the canvas’s HTML architecture.[2]

That placement made the output part of the next prompt. Users could resize a prototype to explore responsive layouts, arrange alternatives side by side, and draw revisions over a result before generating another version. Exported images showed iframe contents as blank rectangles, so tldraw supplied the previous result’s HTML with instructions to fill in the blank region. It also sent canvas text separately to compensate for imperfect image-text recognition. These mechanisms made annotation and iteration central to the workflow.[2][1]

In the interview published January 5, 2024, Ruiz compared four kitchen-timer prompts: a labeled box, a UI drawing, a state chart, and the drawing combined with the state chart. The drawing alone produced buttons that did not work; the combined specification produced a working timer matching the interface more closely. Another result counted upward until Ruiz annotated it to count down. The prototype returned a single HTML file containing markup, styles, and JavaScript, displayed in an iframe; that version did not support React projects or multiple pages.[1]

Ruiz’s November 18, 2023 account describes a hurried hosted launch on Wednesday night. API rate limits immediately broke the application, prompting the team to add an input for users’ own OpenAI developer keys. On Friday, it released a starter template so developers could build their own tldraw/GPT-4V experiments.[2]

With assistance from Fal.ai, tldraw also applied latent consistency models to rapid image-generation experiences. DrawFast placed generated images beside sketches and updated them as users changed the drawings. Lens sampled a square from a vertically scrolling collaborative drawing surface and transformed it into imagery guided by a scene prompt. Ruiz recalled that Lens combined Cloudflare Workers and CRDT-based collaboration with a laptop in his living room performing the screenshot capture. He emphasized the shared experience of interacting with changing imagery rather than preserving a particular output.[1]

Ruiz resisted making these demos the company’s sole direction. In the interview published in January 2024, he argued that a drawing app attached to a model offered little defensibility, while tldraw’s refined, extensible canvas was harder to reproduce. He wanted the SDK to support many experiments and model comparisons—an ambition he described as being the “battlefield for multimodal prompting and multimodal AI experience”.[1]

The canvas underlying those AI experiments predated them. Its drawing tools grew out of Ruiz’s earlier digital-ink work, while demand from developers shaped tldraw into an infrastructure company after its 2021 launch.[1][6]

Ruiz came to software through fine art and product design. He earned a master’s degree in visual art at the University of Chicago, then shifted toward design and technical prototyping at age 30. Using Framer helped him learn to code; he subsequently worked on Framer’s education content and at the design-tool company Play. His open-source projects included Perfect Arrows, an exploration of aesthetically pleasing connections between points or boxes, and Perfect Freehand, which converts input points into an enclosing polygon that can be outlined and filled to produce variable-width vector ink.[1]

After integrating his ink into Excalidraw and building experimental infinite canvases, Ruiz began developing a general-purpose whiteboard engine in spring 2021. He recalls that an unnamed communications company paid $75,000 for source access while adapting a touchscreen whiteboard for browsers. GitHub sponsors also gained early access. He released the app and library under the MIT license in November 2021 and recalled roughly 40,000 launch-day visitors. By the period leading to his seed fundraising, he had received approximately $200,000 in cumulative sponsorship, with recurring sponsorship around $5,000 a month.[1]

Developer demand changed his plans. Ruiz had arranged to join Adobe, but after the release, teams approached him to embed canvases in applications that were not themselves whiteboard products. He declined the planned start, raised funding, and assembled a London team around a developer SDK. React-based customization let those teams put their own interactive components on the canvas instead of building its selection, camera, undo, and geometry systems themselves.[1]

By November 2022, tldraw was developing a full rewrite: both an application and an extensible library for spatial interfaces, accompanied by a custom real-time collaboration engine. The library supplied interactions such as selection, rotation, resizing, snapping, editable text, undo and redo, file persistence, and embedded content. The rewritten version was not yet open-sourced. Ruiz named David Sheldrick, Jaime Blair, Lu Wilson, Alex Dytrych, Faraz Shaikh, Mitja Bezenšek, and Sunil Pai as collaborators in the seed announcement.[3]

On August 5, 2024, tldraw released tldraw sync as the SDK’s first add-on module. Originally built for its own whiteboard, the collaboration engine had, the company reported, served more than 400,000 users across 200,000 shared projects. Production deployments were self-hosted and connected to the developer’s storage; the hosted demo server erased its data after 24 hours. The engine sends essential updates, resolves conflicts, and coordinates with clients to avoid duplication. It handles canvas collaboration alongside an application’s existing data systems.[5]

SDK 4.0, announced September 18, 2025, made production licensing an enforced requirement: production use needed a valid trial, commercial, or hobby license key, while development use required no license. Its starter kits remained MIT licensed and covered canvas agents, workflows, branching chats, image annotation in chat, and multiplayer applications. The announcement credited Sarah Fossheim for accessibility work and reported that the SDK had achieved WCAG 2.2 AA compliance.[7]

tldraw announced a $2.7 million seed round led by Lux Capital on November 30, 2022, with Amplify Partners and individual investors participating. On November 22, 2023, it announced a $2 million insiders’ extension from Amplify Partners, Lux Capital, and Preston-Werner Ventures, all prior seed participants. Ruiz said the extension closed a week before Make Real went viral. A $10 million Series A, led by Lux Capital and Definition, followed on April 9, 2025.[3][4][6]

At its April 2025 Series A announcement, tldraw reported $1 million in first-year sales revenue and just under $1 million in closed license deals. Named SDK applications included ClickUp Whiteboards, Padlet Sandboxes, Jam’s screenshot annotations, and Mobbin’s internal tools. tldraw reported a team of nearly twenty across London, Europe, and the United States, and identified tldraw.computer as a recent collaboration with Google DeepMind.[6]

By September 18, 2025, tldraw reported 40,000 GitHub stars, more than 70,000 weekly installs, 8,000 Discord members, and 2,000 pull requests opened that year in its repository. These measures describe developer activity and community reach; the earlier sync figures describe collaboration-engine usage.[7][5]

On March 3, 2026, tldraw launched its MCP App in Cursor. The server returns an interactive canvas within the chat, exposes tools to create, edit, and delete shapes, and feeds the current canvas state back into the agent’s context. A user can therefore revise a diagram and ask the agent to revise its plan in response. Support for Visual Studio Code, ChatGPT, and Claude was announced as planned.[8]

SDK 5.0, announced May 7, 2026, added reactive themes, detailed display values, and custom geometries. Temporary overlays such as indicators and handles moved from individual React-rendered SVGs to a single HTML canvas. New modules converted Mermaid diagrams into tldraw shapes and provided a Driver for simulating editor behavior, drawing on helpers previously used in the test harness and intended to support MCP and AI features.[9]

The free tldraw offline desktop application launched on July 16, 2026, with local files, no accounts, and no server. Ruiz traced it to a winter-2024 experiment that he shelved, then resumed during winter break 2025. He recalled that coding agents helped replace whole-document saving after every change with updates that were merged into the document. Previous work on Fairies supplied a harness for mediating between agents and the canvas.[10]

Offline opened a local port accepting JavaScript with access to the editor and window, allowing computer-based agents such as Claude Code, Codex, and Pi to manipulate the canvas. It distinguished one-off executions from persistent scripts that run again when a file opens. The new .tldraw format bundles canvas data, assets, and those scripts: a file can contain behavior, such as a drawn bin that deletes shapes dragged into it, as well as a static diagram.[10]

On September 10, 2026, product engineer Max Drake and interns Audrey Leo and Leo Knudsen introduced the free SDK-based animation tool tldraw flash. It records actions on the canvas and lets users reposition animated shapes without recording their movements again. Animations can be exported as MP4s or shared through a gallery for remixing. Flash extends the company’s work into recorded performances using the same canvas mechanics.[11]