D2's Creator Stopped Writing Code a Year Ago. TALA Is Free Now.

TL;DR

On September 5, D2 creator Alexander Wang announced that the for-profit company behind it, Terrastruct, is shutting down and the project is now fiscally sponsored by Hack Club. Two days later, TALA, the proprietary layout engine that funded D2's open-source half, shipped under MPL-2.0 inside D2 v0.9.0.

You now run it with d2 --layout=tala. No plugin install, no license key, no watermark, no Enterprise tier required to put it in CI. The other half of the announcement is the part that will get argued about: Wang says every contribution he makes from here will be written by AI, and that he has no intention of ever putting an LLM inside the tool.

The company is gone, the code is not

D2 was a textbook open-core arrangement. The language and CLI were free and MPL-2.0 on GitHub, currently sitting at about 25,200 stars. The money came from two closed pieces: D2 Studio, the IDE, and TALA, the layout algorithm. Wang's post is blunt about the arithmetic: "The development of open-source D2 was funded by the closed-source IDE (D2 Studio) and proprietary layout algorithm (TALA). The company is shutting down."

What happens next is the unusually good version of this story. Instead of the repo going quiet and the paid binaries going dark, both closed products are being opened. TALA landed first. D2 Studio is promised but not yet shipped, and when it does arrive Wang says it will be purely client-side and usable offline.

He is also explicit that his attention is elsewhere: "I will continue to maintain and improve D2. But my time commitment will be limited. It's not my primary focus anymore." His GitHub profile currently lists him as infra at OpenAI. He is looking for maintainers at [email protected], with paid contributor contracts funded by donations to the non-profit.

If you were paying for TALA, here is what you were paying for and no longer are. The old product page distinguished Personal from Enterprise licenses, and the difference mattered: only an Enterprise license permitted running TALA on servers such as CI/CD, and diagrams made under one had their IP automatically assigned to the purchasing entity, revocable when an employee left. Students, researchers, non-profits and open-source maintainers could email for a discounted license at $5 a month. Unlicensed renders carried a watermark. All of that machinery is now a historical footnote, and the old TALA repo is archived.

What TALA actually is

TALA stands for Terrastruct's AutoLayout Algorithm, and its design goal is narrow on purpose: software architecture diagrams. It is primarily an orthogonal layout engine rather than a DAG engine, which is Wang's way of saying it tries to produce the thing you would have drawn on a whiteboard instead of a tree that grows relentlessly in one direction.

Under the hood it scores candidate layouts against several aesthetic objectives at once, including symmetry, median distance, flow, and clustering of like nodes, blending published graph-drawing techniques (cited in the source, which you can now read) with original ones. Think of it less as a solver with one right answer and more as a picky art director trying three arrangements of the same furniture and keeping the one that looks least like a server rack.

The tradeoffs are real, and the post says so

To Wang's credit, the release post leads with limitations rather than burying them:

• It is stochastic. TALA runs 3 seeds by default and keeps the best-scoring result. Same seeds plus same input gives you the same diagram, but add a single node and the whole layout can rearrange. Dagre and ELK mostly keep their shape and slot the new node in. For a diagram in version control that reviewers diff, that is not a small difference. You can pin it with --tala-seeds or D2_TALA_SEEDS.
• It is worse at DAGs. Wang says he still reaches for Dagre or ELK when he wants a long flowing graph.
• It scales nonlinearly. Bigger diagrams cost disproportionately more time.

That last point has numbers behind it. The project published d2-benchmarks on September 7, a reproducible harness that times the full CLI round trip (start, lay out, render, write, exit) against Mermaid, Graphviz, and PlantUML, 20 runs after 3 warm-ups on an Apple M4.

Read that chart honestly. TALA is the expensive layout in D2, not the fast one. It is also the prettiest one, and the whole D2 binary with TALA attached is still several times quicker than the tools most teams currently paste into a wiki. The benchmark README is careful to note the numbers come from one desktop session with uncontrolled background activity, so treat them as directional rather than lab-grade.

Why an agent should care about a layout engine

Here is the part that makes this a 2026 story rather than a 2022 one. TALA lets you lock node coordinates with top and left, and mix pinned nodes with automatic ones in the same diagram. Wang's framing: models "can draw in 2D space well, but TALA still takes care of routing, which models still struggle with." The example diagrams in the release post for that section were themselves AI generated.

Wang's stated roadmap goal is to "be the most powerful harness for agents," and his reasoning is the most useful paragraph in either post. Models are good at almost-perfect diagrams, he writes, but the output is "too non-deterministic, uncustomizable, imprecise," so "you still want some text-based source to back the artifact. Kind of like HTML for web pages instead of drawing every pixel." If you have ever asked a model for an architecture diagram and received a beautiful PNG you could not edit one box of, that sentence is the whole argument.

v0.9.0 is also just a much better release

Lost in the licensing news: v0.9.0 is the biggest performance jump D2 has had. The built-in renderer now exports PNG, GIF, PDF and PPTX directly, Markdown labels render as native SVG instead of HTML foreignObject content, and the render path got substantially cheaper.

In the release notes' own words: SVG rendering is approximately 10x faster across the end-to-end corpus, real-world diagrams compile, lay out and render approximately 3x faster at the median, and SVG exports come out approximately 24% smaller across that corpus and 18% smaller on real-world fixtures, with unchanged appearance. For anyone regenerating a diagram set in CI on every merge, that is the line item that shows up on the bill.

The part that will start arguments

Wang did not hide the AI policy in a footnote. "D2 thus far has been a product of handcrafted code. That era is over. This next phase of D2 will be driven by AI. By this, I mean that all of my contributions will use AI. I welcome AI contributions and will use AI to review your AI, etc." The one carve-out is prose: he says he will author all blog posts and comms himself.

The credential he offers is a striking one. "I wrote code ~every day for 10+ years, and have not written any in the past year even though I'm working with software every day." He acknowledges the choice is controversial and links to the downsides himself, which is more intellectual honesty than most AI-workflow announcements manage.

And then, in the same post, he draws a hard line in the other direction. "I still don't see a place for LLMs integrated into diagramming tools. I think models are smart enough to just look at some examples and then take your specs or code or whatever you want and turn it into a .d2 file. There doesn't need to be yet another thinly tuned LLM interface for D2 to provide."

Non-profit status enforces the rest of the discipline. No MCP server, no API, no user accounts, no server rendering, no multiplayer, because none of those can exist without infrastructure someone has to pay for. The scope is also staying narrow on principle: software architecture diagrams, and no Venn diagrams even though AI would make them trivial to add.

Write every line with a model, refuse to ship a chat box. In a year where the reflex is to bolt an assistant onto anything with a text field, that is the more interesting half of the position.

What to do with this

Practical, in order:

• Upgrade to v0.9.0 and try d2 --layout=tala on a diagram you already have. It is one flag and the comparison is immediate.
• If you have TALA in CI on an Enterprise license, that line item is now dead. Same for any watermark workaround.
• If your diagrams live in git and get reviewed, pin D2_TALA_SEEDS or stay on Dagre. The stochastic relayout is a genuine review hazard, not a theoretical one.
• If you are generating diagrams with agents, the pinning path is the actual news here: let the model choose coordinates, let TALA route.
• The playground runs fully client-side, so you can evaluate it without installing anything.

The one open risk is continuity, and it is the same risk it always was, just newly visible. A 25,000-star project now depends on a maintainer who has said it is not his primary focus, working through models, funded by donations, looking for help. That is a more honest description of a lot of your dependency tree than most maintainers put in writing.

Key Takeaways

• TALA is MPL-2.0 and bundled. D2 v0.9.0 shipped September 7 with d2 --layout=tala, no plugin, no license key, no watermark, and no Enterprise tier needed for CI.
• The open-core business is over. Terrastruct is shutting down; D2 continues as a non-profit fiscally sponsored by Hack Club, with D2 Studio promised as client-side and offline.
• TALA is the pretty option, not the fast one. On the project's own benchmark it takes about 4x Dagre's time on a real-world diagram, is weaker on DAGs, and scales nonlinearly.
• Its randomness is a review hazard. Three seeds by default and best-score wins, so one added node can rearrange the whole diagram unless you pin seeds.
• The agent story is coordinate pinning. Models place nodes well and route connections badly, so TALA's hybrid pinned plus automatic mode splits the work along that seam.
• The maintainer will write all future code with AI and no LLM will ship inside the tool. Text-based source stays the artifact of record, which is the same argument as HTML over pixels.

Sources: D2 is non-profit (Alexander Wang, Sept 5, 2026), TALA is open-source (Sept 7, 2026), D2 v0.9.0 release notes, d2lang/d2-benchmarks, terrastruct/TALA (archived), terrastruct.com/tala, archived February 2026 (licensing terms), d2lang/d2 on GitHub