Seven Adobe apps, rebuilt by AI: the file formats are the real story
What shipped
In the last days of September, the ArtCraft studio published seven open-source applications that recreate Adobe's most-used creative tools: PhotoCraft for image editing, VectorCraft for vector illustration, FilmCraft for video, LightCraft for photography, and equivalents of Acrobat, After Effects and InDesign. Pure Rust, dual-licensed MIT or Apache-2.0, native apps for macOS, Windows and Linux, with browser builds on top. The first repository commits are dated September 30 and October 1, and the developer, Brandon Thomas, says the work is driven largely by AI agents running Anthropic's Claude Opus 5.5. Five more apps have arrived since, covering word processing, audio, slides, CAD and spreadsheets.
The applications are clean-room reimplementations: written from public specifications and observed behavior, with no Adobe code, shaders, ICC profiles or assets allowed in, per the rulebook each repository carries. That rulebook, an AGENTS.md file, opens with a priority list whose first item is "never crash". Strange until you consider what is rebuilt: in an editor, a crash means lost work, and an implementation that never loses work is worth more than one that adds features faster.
How the agents worked
Each agent got one subsystem: the PSD parser, the GPU compositor, the text engine, the filter set. A human did the integration, the testing and the releases. The numbers in the repositories are the project's own claims, so read them that way: PhotoCraft lists more than 1,700 tests and says that re-saving keeps the render of 307 of the 309 files in a public PSD test set. EffectCraft has a first commit dated October 1 and claims 306 effects, which its README says covers all 298 in After Effects.
The attention was real. Together the repositories passed 79,000 stars, and PhotoCraft alone sits above 45,000.
Where it is honestly weak
Almost everywhere that matters for paid work today. The project labels the apps early alpha, and the READMEs agree with the label. EffectCraft cannot open After Effects projects and Adobe plugins do not run on it. LightCraft reads some raw formats but not Nikon's compressed files or Canon's, a gap the team attributes to its own clean-room rules. Cloud libraries, shared assets and the plugin ecosystem do not carry over. By the project's own numbers, functional parity for common workflows sits around 86 percent, and the share of professional work that could move over today is between 25 and 35 percent.
The moat was never the menu bar
Adobe sells tools, but what keeps people is formats. A layered PSD is the reason a designer does not leave: nothing else opens it reliably, so the files stay, and so do the people holding them. That is why the quietest claim in this project is also the loudest. A proprietary format holds as long as documenting it is expensive. Agents just made documenting cheap, and a free alpha that round-trips almost every test PSD is what that looks like two weeks in.
Follow the money from there. If a feature list that took decades to assemble can be approximated in weeks, then anyone selling software whose value is mostly a feature list should assume a credible clone is coming. What resists is what cannot be rebuilt from public specifications: the ecosystem around the format, the habits built on top of it, the services, the legal guarantees. Adobe's own answer this year has been to wire its tools into AI assistants. These replacements are built the other way around: no account, no subscription, every menu item exposed to an agent as a command.
The legal question nobody has settled
Clean-room reimplementation is old and legally well-trodden. One team documents how a system behaves; another writes new code from that description. The BIOS clones of the 1980s were built this way. What is new here is that agents played every role in the room, and that unsettles the doctrine. A lawyer publicly described the project as "AI-assisted decompilation", which would contradict the clean room; nobody has shown decompiled code inside the repositories, and the project's documentation says the opposite. Ars Technica raised a separate risk: mimic the interface closely enough and "trade dress" becomes a problem, wherever the code came from. Underneath all of it sits the question no court has tested: the models learned from public code, some of it probably Adobe's, and nobody knows yet how much that weighs.
What I take from it
If you pay for these tools, nothing here says cancel this quarter. The apps are alpha by their own admission, and the ecosystem studios actually depend on is not in the box.
If you sell software, the lesson generalizes past Adobe. Assume the documented surface of your product can be reproduced in weeks by someone with agents and patience, and spend accordingly: on the parts that are not a feature list. And if you write code for a living, watch the shape of this project more than the product. An agent does not need to be right about everything. It needs to get to try everything, for hours, against tests that decide what survives. That pattern is the real announcement here.
Need a hand with a WordPress site that keeps breaking? See what I do and how I work.