The Adobe clones went viral—then hit reality
AI agents just made software copyable—the hard way. Take ArcCraft, the open source suite from Brandon Thomas Thomas, which went viral for its audacious clean-room claim: PhotoCraft, VectorCraft, and FilmCraft were built from public specifications and observed behavior, not Adobe’s proprietary code or assets. The premise was striking.
Hands-on experience, however, brought a dose of reality. A Hacker News commenter, with experience in design and film, reported that basic layout projects simply did not work. PhotoCraft’s own “honesty parity assessment” on its roadmap confirms Photoshop parity remains well below 50%.
This distinction between a familiar-looking interface and a dependable creative tool is critical. Production workflows rely on seamless file compatibility, robust handling of edge cases, predictable rendering behavior, and years of accumulated polish. ArcCraft, despite its viral moment, highlights that true software replication means more than just a matching UI; it demands a deep, functional parity that current AI-assisted clean-room efforts have yet to achieve.
An MCP server gives agents a view inside binaries
ArcCraft’s clean-room claim, however tenuous, points to the underlying technical shift: AI agents now reverse engineer binaries with unprecedented precision. The number one trending repo on GitHub, Reverse Engineer Anything (REA), isn't magic; it’s an MCP (Model Context Protocol) server that bridges AI agents to professional reverse-engineering tools like Hopper, Ghidra, and IDA. This setup gives agents a programmatic view deep inside compiled software.
This isn't one-click cloning. Agents use REA to inspect functions, analyze pseudocode, and infer data structures, just like human engineers. They then apply these findings to guide the reconstruction of source code, distinguishing analysis assistance from instant replication. It’s a sophisticated process of architectural inference, not illicit code extraction.
MCP provides a standard way for agents to interact with external tools, moving beyond mere text prompts. However, real-world constraints persist: setting up the MCP server, securing access to licensed reverse-engineering software, and robust human verification remain critical. The agents perform the grunt work, but a human still directs the orchestra and validates the output.
A 1990s pinball game became a Rust port
A far more compelling demonstration of agent-driven reverse engineering came from the Better Stack video, which detailed an experiment to port the classic Windows game 3D Pinball: Space Cadet to macOS. This wasn't a clean-room rewrite from public specs; it was a direct assault on a legacy binary.
Claude Opus 5.5, the star of this show, took a downloaded Windows executable from the Internet Archive and, in a reported 2 hours and 25 minutes, generated a native macOS port. The process involved orchestrating 15 sub-agents to tackle distinct game components: physics, flippers, and all 32 missions.
The results, according to the video, were staggering: 31,000 lines of Rust, 174 passing tests, and zero external dependencies. This outcome suggests agents can accelerate complex legacy porting tasks for bounded applications.
But let's be clear: while impressive, these metrics do not guarantee feature-for-feature fidelity, long-term maintainability, or legal clearance. Code volume and passing tests merely indicate functional parity within the agent's interpreted scope, not necessarily against the original game’s intricate behavior or intellectual property. For more on the underlying technology, see GitHub - morluto/rea: Reverse Engineer Anything.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
Closed code is a weaker moat, not an open-source switch
Claiming that "everything is open source" now overstates what AI agents can reliably reproduce, even as they make partial recreation cheaper. While ArcCraft's PhotoCraft, VectorCraft, and FilmCraft highlight the potential of clean-room rewrites from observed behavior and public specifications, the project’s own roadmap admits Photoshop parity remains “well below 50%.” User reports of "joke software" for basic layout projects underscore the gap between concept and production-ready code.
Technical and legal limits persist. Complex proprietary formats, deep cloud integrations, and vast datasets resist imitation, especially when critical edge cases define usability. Clean-room claims, while a legal defense, do not automatically resolve copyright, licensing, or anti-circumvention concerns, particularly when a direct competitor emerges from automated reverse engineering.
Defensibility will shift from code obscurity toward data, infrastructure, network effects, and tightly integrated ecosystems. While AI agents like Claude Opus 5.5 can port 32-bit Windows EXEs to 31,000 lines of native Rust in 2.5 hours, these are impressive experiments, not tested drop-in replacements for real work. Until such clones prove their mettle in demanding professional workflows, treat them as fascinating proof-of-concept, not a signal to abandon your existing software.
Frequently Asked Questions
Is PhotoCraft a full Photoshop replacement?
No. Its roadmap acknowledges that Photoshop feature parity is well below 50%, and users have reported basic project shortcomings.
What does Reverse Engineer Anything do?
It connects AI agents to tools such as Ghidra, Hopper, and IDA so they can inspect and analyze compiled software.
Can AI legally clone any software?
No. Reverse engineering and clean-room reimplementation can raise copyright, licensing, and anti-circumvention issues. The law depends on the facts and jurisdiction.
What did the Space Cadet Pinball experiment demonstrate?
An AI agent reportedly used 15 sub-agents to recreate the game in Rust, producing a native build with 174 tests in about two hours and 25 minutes.