In Germany and the EU, the law explicitly allows you to take apart a product you bought freely in order to find out how it works. Section 3(1) no. 2(b) of the German Trade Secrets Act and Article 3(1)(b) of the Trade Secrets Directive (EU) 2016/943 say so. Since early October, there have been free rebuilds of Photoshop, Premiere and Word. Written by an AI.
That raises two questions. How does an AI agent rebuild in days what vendors have worked on for decades? And is it theft? One thing up front: according to their founder, these rebuilds were created without looking at the code of the originals and therefore also without using that code. This article assesses the technology, the law and the consequences for companies that buy or build software.
What happened
All of the rebuilds come from a single project: ArtCraft, founded by Brandon Thomas from Atlanta, formerly CEO of Storyteller.ai and before that an engineer at Square. The code lives on GitHub under the organization storytold. The first commit of PhotoCraft dates from 30 September 2026. On 3 October, Thomas announced an entire Adobe suite on Reddit, as PetaPixel reported. On 7 October, WordCraft, GridCraft and DeckCraft followed for word processing, spreadsheets and presentations (The Register).
Today the project page lists twelve apps. Besides PhotoCraft for image editing, these are VectorCraft for vector illustration, FilmCraft for video editing, LightCraft for photo library and raw development, PdfCraft for PDF, EffectCraft for motion graphics, DesignCraft for layout and publishing, WordCraft, SoundCraft as an audio workstation, DeckCraft, CADCraft and GridCraft. All are written in Rust, open source under MIT and Apache 2.0, and run offline on macOS, Windows and Linux as well as in the browser via WebAssembly. All are in alpha.
PhotoCraft version 0.7.0 handles layers, masks, adjustment layers and layer styles. Rendering runs natively on the Metal, Vulkan, DirectX 12 and WebGPU graphics interfaces, without Electron. The program opens and saves PSD and PSB files. According to the project page, 134 of 135 real test files are saved back byte for byte. WordCraft version 0.5.0 offers a ribbon, styles, tables, track changes, bibliographies and mail merge. It reads and writes .docx natively based on the public ECMA-376 standard. According to the project page, its 389 commands can be controlled through the ribbon, a command line, a JSON interface and an MCP server. An AI agent can therefore operate the word processor directly.
The project documentation is more cautious about maturity than some headlines. On 5 October, the PhotoCraft roadmap estimated its readiness for professional use at 25 to 35 percent. The WordCraft README cited around 60 percent on the way to a full Word replacement. A pixel artist told Decrypt that PhotoCraft was completely unusable. Thomas, who had told Ars Technica that “software is over”, revised his promise of full parity within a month to months rather than years. Adobe declined to comment, and Microsoft only confirmed receipt of the request. No lawsuits are known so far.
What matters is how it was built. On Hacker News, Thomas wrote that the project uses plain Claude Code and is especially careful not to touch the reverse engineering tool Ghidra. Adobe sources must not be looked at. These rules are also in the code. The AGENTS.md in the WordCraft repository forbids the AI to read, disassemble or copy anything in the Word program package. Word may only be observed from the outside. The AGENTS.md of PhotoCraft states that Photoshop was only studied for behavior and appearance and that code, shaders, profiles and assets must never be copied.
How do you build something like this? Two paths
In parallel with ArtCraft, a tool went viral in the same days that takes a different path. REA, short for Reverse Engineer Anything, is an open-source MCP server that launched in April 2026 as “betterBinaryMCP”. It connects AI coding agents such as Claude Code to established reverse engineering tools: Ghidra, which the NSA released in 2019, Hopper, IDA and Chrome DevTools for web and Electron apps. According to Decrypt, REA went from under 10,000 GitHub stars on 6 October to over 100,000 on 12 October, more than 30,000 of them on a single day.
The process, as MindStudio describes it: you start with a plain-language instruction. The agent has the target program decompiled and receives pseudocode with evidence and a list of open questions. From this it writes its own implementation, tests it against the original and repeats the process feature by feature. The REA README is restrained. It claims neither to recover original source code nor to clone applications automatically.
According to its own statements, ArtCraft deliberately did not take this path. Decrypt and Kingy.ai found no evidence of REA being used in independent reviews. Both stories still show the same pattern: an AI agent does the legwork of rebuilding, while humans set the rules and check the results.
| Path 1: black box and specification | Path 2: looking inside | |
|---|---|---|
| Example | ArtCraft, according to its own statements | what REA enables |
| Source of information | menus, keyboard shortcuts, behavior, screenshots, public format specifications such as Adobe’s PSD specification and ECMA-376 | decompiled program code as pseudocode |
| Control | fixed rules in an AGENTS.md on what the agent may and may not do | plain-language instruction, comparison with the original |
| Progress | test corpora, for example whether real files are saved back unchanged | tests against the original, feature by feature |
| Legal position | the established clean room principle | decompilation in the EU only allowed under very strict conditions to achieve interoperability |
At its core, path 1 is the old clean room principle, except that the programmer is an AI. Path 2 is faster at understanding details, but legally much more delicate.
MindStudio tested how much looking inside actually helps. The Mac app Preview was rebuilt once with and once without REA. Both versions worked. The version without REA, based only on screenshots and the agent’s own reasoning, was even finished first. MindStudio therefore calls REA a connector, not a new technology, and the hype partly memetic. The real development, it says, is that software is becoming a commodity and thin wrapper apps are losing their defensibility.
The spectacular path is not necessarily the more successful one. A regular coding agent with screenshots and a specification got just as far in the test.
Reverse engineering is older than the PC
Reverse engineering means analyzing a finished product backwards until you understand how it works. It is old and has often been legitimate competition.
The best-known example dates from 1984. Phoenix Technologies rebuilt the BIOS of the IBM PC using a clean room process. One team analyzed the original and wrote specifications. A separate developer wrote the new implementation from them. He was familiar with other processors and was chosen precisely because he had never seen the IBM code. The process paved the way for the PC clones.
Later examples include Samba for Windows file sharing on Linux, Wine for Windows programs on other operating systems and the Office compatibility of LibreOffice.
So what is new with ArtCraft is not the principle. What is new is who takes on the role of the clean developer. In 1984 it was a person who had never seen the original code. In 2026 it is a language model that a text file tells what it must not look at. Whether that is equivalent is one of the open questions of this story.
Is it theft?
The short answer: rebuilding functionality is generally permitted in the EU, copying code is not. The line runs where someone looks inside a program or where there is unfair imitation. For the latter, additional unfair acts must be present (deceiving buyers, exploiting reputation or dishonestly obtaining the knowledge or documents needed for the rebuild, Section 4 no. 3 of the German Act against Unfair Competition, UWG). The following assessment does not replace legal advice in individual cases.
The foundation is the CJEU ruling in case C-406/10, SAS Institute v World Programming, of 2 May 2012. The court ruled that the functionality of a computer program, the programming language and the data file format are not forms of expression protected by copyright (analysis on netzpolitik.org, in German). The specific code is protected. One important detail: World Programming had no access to the source code and did not decompile.
German copyright law draws the line in two sections. Under Section 69d(3) of the Copyright Act (UrhG), a lawful user may observe, study and test a program to determine the ideas underlying it. Under Section 69e UrhG, decompilation without consent is only permitted if it is indispensable to achieve the interoperability of an independently created program. The information obtained must not be used for a program substantially similar in its expression. This is exactly where the line lies for a rebuild based on decompilation.
Section 3(1) no. 2 of the Trade Secrets Act (GeschGehG) adds to this. Since 2019, it has explicitly permitted reverse engineering of publicly available products. Patents, designs, trademarks and contracts set the limits. These risks exist regardless of the code. Anyone who copies the look and overall appearance of a product too closely, uses third-party trademarks or infringes patents has a problem, even if not a single line was copied. Whether bans on reverse engineering in license terms are enforceable is disputed.
Vendors therefore often try to prohibit reverse engineering in their terms of use. For such a (contractual) prohibition to apply, however, you must be a contracting party and have agreed to the terms of use.
In the US, the line is similar but justified differently. In Sega v Accolade, a court ruled in 1992 that disassembly can be fair use. In 2021, the Supreme Court decided in Google v Oracle that copying API declarations was fair use.
One question remains that the law has not yet answered. A classic clean room separates two human teams. A language model, by contrast, was trained on vast amounts of public code. Whether such a model can be “clean” at all is unresolved. A first debate arose in March 2026, when the Python library chardet was rewritten with Claude and a dispute over its relicensing followed. Lawyer Neil Brown of decoded.legal considers independent reimplementations defensible, but expects litigation as soon as opponents with deep pockets are affected. Thomas also expects pushback, but for the other path. In his view, using REA on such a prominent project would likely result in lawsuits.
What this means for your company
Software used to be expensive to build and cheap to copy. Now building is getting cheap as well.
For software vendors, this means features alone are no longer a moat. If an agent can rebuild the interface, commands and file format as an alpha within weeks, the feature set no longer protects you. What remains are data, deep integration into processes, support, certifications, liability and trust.
For software buyers, negotiating power grows and the number of alternatives increases. A credible open-source rebuild changes the starting position for the next license renewal. But alpha software is no substitute for productive core processes. Maintenance, security updates, compliance and the question of who is liable remain the deciding criteria. A repository on GitHub answers none of them.
For in-house development, make-or-buy calculations shift because building with AI agents is getting cheaper. This applies less to the next Photoshop than to internal tools and specialist applications for which a license was bought because building them was too expensive.
The other side belongs in an honest assessment. Rebuilding an interface is not rebuilding an ecosystem. Nobody catches up on decades of compatibility, plug-ins and edge cases within weeks. The one test file out of 135 that PhotoCraft does not write back byte for byte stands for exactly these edge cases.
Conclusion: three things you should do now
Check your moat. If you sell software yourself, ask for each product what a competitor with an AI agent could not rebuild within three months. If the answer consists only of features, that is the task for your next strategy round. Include a prohibition of reverse engineering in your terms of use. Enforcing it, however, is legally difficult.
Recalculate make-or-buy. For upcoming license renewals and internal tools, evaluate the build option with realistic costs for agent-assisted development. Factor in operations, security updates and liability in full.
Set rules for coding agents. If your teams work with AI harnesses, define in writing what an agent may and may not look at. An AGENTS.md with clear prohibitions, as ArtCraft uses, is a simple start. Decompiling third-party software does not belong in the toolbox without a legal review.
If you want to reassess your AI strategy or a specific make-or-buy decision against this background, we are happy to talk it through.