the idea
canvas-as-infrastructure — the lineage tldraw inherited, what it proved, and what agents do to it.
lineage
- 1963sketchpadsutherland's sketchpad — direct manipulation of geometry as computation, the ancestor every canvas interface cites. tldraw inherits the claim that drawing on a surface is a way of programming.
- 1970sthe dynabookkay's dynabook — the personal dynamic medium. the canvas as a place where a person builds their own tools rather than consuming someone else's documents.
- 1987hypercardhypercard — end-user authoring with a real object model underneath. the direct ancestor of 'a canvas is a thing you build with': the surface is friendly, the substrate is programmable.
- 2019–2020spatial softwarejohn palmer's spatial interfaces and spatial software essays — 'a world within its interface' — the 2019–2020 genre that made the spatial canvas a legible product category again before the tools caught up.
- 2019–2023musemuse — the thinking-tool canvas that wound down in fall 2023. adam wiggins wrote the post-mortem and sits on tldraw's cap table: the predecessor's failure is literally inside the successor's investor list.
- 2020–excalidrawexcalidraw — the mit-licensed hand-drawn canvas that adopted perfect-freehand. the open alternative that defined the aesthetic and now defines the license comparison tldraw keeps losing.
- 2024–the canvas revivalthe zoomable-ui re-emergence — artifacts, chatgpt canvas, mistral's le chat: the frontier labs made 'canvas' an ai-interface paradigm while tldraw was already there. the category arrived around the company.
what tldraw proved
- infrastructure beats the appcanvas-as-infrastructure survives where canvas-as-app struggles — ~$1m of first-year license revenue, a named roster (clickup whiteboards, padlet, autodesk, replit, shopify, luma, runway, blackrock), and a $10m series a on the strength of it. muse wound down; tldraw compounded.
- the dom betthe html-over-webgl bet paid — 'i could probably just make normal website stuff just as fast… then i would be able to put anything onto that canvas.' a dom canvas can host iframes, which is what made make real possible at all; ruiz argues miro 'could not' have built it.
- demos as r&ddemos are the growth engine — make real, draw fast, computer, fairies each generated the largest coverage moments in the record, and each funnels into sdk awareness. ruiz calls the program 'marketing, r&d, and fun'; the hn numbers back the marketing part.
- permanence as trustpermanence is a feature — v1 parked at old.tldraw.com 'until the end of time' and released under mit is the trust primitive that makes the license argument survivable. the company keeps an older self alive on purpose.
- the agent surfacethe canvas is a real agent surface — carry on, teach, computer, the agent starter kit, fairies, and mcp access all ran on the same substrate. 'if it's busy, you just go somewhere else' is a multi-agent argument a chat window can't make.
what stayed unfinished
- the app is a showroomthe app never became the business — tldraw.com stayed free, the sync demo server wipes data after 24 hours, and offline is a free desktop app. the product-shaped canvases are showrooms for the sdk, not revenue.
- the spent goodwillthe open-source goodwill was spent — 'open source is a spectrum,' the readme correction, the formerly-foss listings, the license-key gate. the community that made the 2021 launch warm now writes the comparisons that recolor tldraw as closed.
- the sunset patterndemos sunset faster than they convert — draw fast went down in january 2024, computer stayed closed source, fairies ran one december. the attention keeps arriving; the products it pointed at keep expiring.
- the pricing opacitypricing stays opaque — the protest issue, the 100-day trial, hobby and commercial paths that still route through sales. 'just ask for money' is easier to say than to publish a price for.
- the cadence nobody watchesrelease news stopped being news — sdk 4.0's thread was a fight, sdk 5.0's was a shrug (1 point). hn attention now arrives only for demos and controversies, not for the product cadence.
the agent-era turn
- forthe canvas is the right agent surface — spatial, parallel, inspectable. 'if it's busy, you just go somewhere else' generalizes to agents: many workers on one visible substrate, which is what the fairy experiments were actually testing.
- forinfrastructure position compounds — the sdk already lives inside clickup, replit, luma, and google's own stitch experiments, so agent features ride existing deployments rather than fighting for new surface area.
- againstthe non-determinism problem is real — natural-language node graphs fail unpredictably, the computer's own hn thread collected failure reports, and the agents are 'essentially working blind.' the canvas can host agents; that doesn't make them reliable.
- againstthe license gate collides with agent-era distribution — a keyed production sdk can't ride inside downstream open source, and the community that builds agent demos is exactly the one priced out. the ai-slop flood is the same ecosystem arriving uninvited.
- nuancethe paradigm is commoditizing — artifacts, chatgpt canvas, le chat, google's own canvases. tldraw's edge is the substrate (dom, sync, licensing), not the concept; if 'canvas' becomes a feature every chatbot ships, the sdk is either the picks-and-shovels winner or a commodity input.
- nuancegating may be the honest trade — 'fate and capital both demand that tldraw be a sustainable project.' if the sdk is infrastructure, someone has to pay for it, and the mit alternative (excalidraw) exists for everyone the license excludes.
successors
- excalidrawthe mit control group — the canvas that adopted perfect-freehand and now anchors every 'fair open-source pick' comparison tldraw loses.
- draw-a-uithe community seed — hood's mit-licensed sketch-to-html demo that tldraw forked into make real. the ecosystem makes the demos; the company packages them.
- musethe predecessor's ghost — the thinking-tool canvas whose post-mortem is the thesis's control case, and whose founder backs tldraw.
- the canvas cohortkosmik and the app-shaped canvases that kept launching and sunsetting around the sdk — the field the infrastructure bet was made against.
- the agent-canvas categorypaper, replit agent 4, google stitch — the aiewf 'spatial harness' talk already treats the agent canvas as a category with tldraw inside it, not above it.
- the lab canvasesartifacts, chatgpt canvas, mistral's le chat — the frontier-lab versions of the same surface, owned by the model providers tldraw demos against.
open questions
- what is tldraw's exact public launch day — the hn thread is november 18, 2021, but the reports leave the precise opening-day sequence unresolved.
- does license revenue compound past ~$1m a year — and does the $10m series a price a platform or a services-adjacent sdk?
- is the linkedin-aggregated '$24.1m total funding' figure a phantom row, or does a real unannounced round sit between the extension and the series a?
- can a production-keyed sdk participate in the agent era it demos for, when the agent ecosystem builds downstream open source the license excludes?
- do the demos ever become products — or is 'marketing, r&d, and fun' the permanent terminal state of make real, computer, and fairies?
- does 'source available' hold as the settled self-description, or does the spectrum framing return each time an llm misstates the license?
- who owns the agent-canvas category — the sdk that hosts the surface, or the labs whose models fill it?
- does tldraw itself become 'a vibe coded alternative to tldraw' — ruiz's own joke about where the tests-to-closed-source bit was actually pointing?
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.