REA AI Reverse-Engineering Tool Exceeds 290k GitHub Stars, Tests Show Limits
The open-source project REA, Reverse Engineer Anything, rapidly accumulated more than 290,000 GitHub stars by October 17, drawing attention for its claim that AI can analyze closed-source software and recreate features without source code. Demos covering Notion’s clipboard handling, DX-Ball’s sound algorithms and Touhou game bullet patterns helped fuel the hype, even prompting tech commentator Matthew Berman to declare software finished. REA did not invent new decompilation technology. It uses the Model Context Protocol to connect established reverse-engineering tools such as Ghidra with AI coding agents including Claude Code and Codex. Instead of manually inspecting disassembled functions, developers can ask an AI to query the tools, interpret results and attempt to write replacement code. Because the workflow arrives as AI agents improve, it has attracted large communities, including a Reddit recommendation post with about 14,700 upvotes. Early projects showed concrete uses. One developer combined Codex, REA and Claude Code to rebuild parts of a Claude Slides editor. Another analyzed a smart toilet’s Bluetooth protocol through a manufacturer’s mini-program and built a menu-bar remote, completing the tool in roughly 30 minutes. The open-claude-ui project documented dozens of interface components and interaction behaviors from Claude’s desktop app. Still, successes have mostly focused on specific modules rather than complete commercial applications. Hands-on tests soon exposed limits. In a GitHub comparison against bzip2, REA correctly traced a target function and its callers and callees. But when testing an OpenMW esmtool binary, it missed existing call relationships in 22 of 50 functions while still reporting the analysis as complete. Other tests found local AI models sometimes ignored REA entirely, even attempting to read binaries as text. Project author Morluto said the original prototype MCP design may not suit most agents as clients evolve, while arguing programmatic use of professional tools remains promising. The broader AI software-rebuilding wave has created additional concerns. Software engineer Brandon Thomas, previously at Square, released ArtCraft, a suite of open-source applications modeled on Adobe products, plus tools inspired by Microsoft Office. Thomas described the approach as clean-room reimplementation rather than using REA or decompiling proprietary code, but editorial testing found significant gaps in automatic selection, image processing and 4K video performance. Reddit debates have also questioned whether AI-assisted analysis and code generation can satisfy independence requirements associated with clean-room development. Downstream effects are emerging. Decomp.dev stopped publishing new project indexes, citing a wave of unverified AI-generated decompilation claims and weakened community quality controls. A counterfeit REA repository was also flagged for distributing a Windows installer marked as malware by Microsoft Defender and multiple VirusTotal engines, though infection counts were unclear. The episode suggests AI has lowered the barrier for some reverse-engineering research and accelerated creation of approximate substitutes, but it has not replaced professional tooling or proven that major software can be automatically restored. The claim that software is done remains unsupported by current evidence.
