REA lets AI agents take apart any app on your machine. Code you ship to customers is no longer a secret.

A free tool called REA, short for Reverse Engineer Anything, has become the most talked-about project on GitHub. Published by a developer who goes by morluto, REA connects AI coding agents such as Anthropic’s Claude Code, OpenAI’s Codex and Cursor to the professional decompilers Ghidra, IDA and Hopper. Point an agent at a compiled program and it can explain how a feature works, show the evidence and rebuild the logic in your own code. It had fewer than 10,000 stars on 6 October and more than 110,000 by 12 October, GeekPark reported from Star History data. By the evening of 12 October the repository page showed more than 140,000.

Stars measure attention, and plenty of them come from people who will never install the tool. The attention is still the story: in one week a large part of the software industry looked at what happens when reading someone else’s compiled code stops requiring a specialist.

What REA does and what it does not do

REA has no decompiler of its own. It runs as an MCP server, the standard plug-in format for AI agents, and wraps existing analysis tools so an agent can call them, AICoder notes. Version 4.1, released on 6 October, works with twelve agents and covers Windows, Mac and Linux programs, Electron desktop apps such as Notion, .NET software, Android apps, websites and firmware. The analysis runs on the user’s own machine.

The examples in the repository are small and specific. One recovers how Notion’s desktop app passes data to the clipboard. Another rebuilds the sound logic of the 1990s game DX-Ball in C, and one module passes 3,205 test cases taken from the original program. The README says plainly that REA does not recover original source code and does not clone whole applications, and that it is meant for lawful research, interoperability, preservation and software the user owns or may examine.

So nobody is pressing a button and getting a free copy of Microsoft Office. What has changed is the cost of the first step. Mapping how a closed program works used to take an expert days or weeks with a disassembler. An agent now does a first pass in an afternoon.

Why reverse engineers call it hype

Experienced reverse engineers are less impressed than GitHub is. Similar connectors such as GhidraMCP and IDA Pro MCP existed well before REA, and Elias Bachaalany, a co-author of the book Practical Reverse Engineering, answered a promotional post with a single word, according to GeekPark: “Hype”. Duncan Ogilvie, who wrote the IDA Pro MCP plugin, said adding more tools does not make the models better.

The technical limits are real too. Security researchers quoted in the same report say current models handle plain or lightly protected code well and stall on heavily obfuscated software. REA has no live debugging. Users complain that its tool descriptions fill the model’s context window. And Rafael Roman, who worked in financial crime prevention at the German bank N26, makes a sharper point: reading a banking app tells you how the app talks to the bank, but the money sits on servers the app cannot command.

All of that is fair, and none of it changes the direction. Coding agents have become good enough, and cheap enough after this autumn’s price cuts, to make reverse engineering a routine task. REA is simply the project that made the industry notice.

What EU law allows when an agent reads your software

European law already permits more reverse engineering than many software vendors assume, and less than many REA users assume. The EU Software Directive lets a lawful user decompile a program when it is needed to make an independently written program work with it. Article 6 also says the information gained may not be used to build a substantially similar program. In 2021 the EU Court of Justice added that a lawful user may decompile software to fix errors that stop it working, and that a licence cannot rule that out entirely.

Trade secret law points the same way. Under the EU Trade Secrets Directive, studying or taking apart a product you lawfully own is a legitimate way to learn how it works, unless a contract says otherwise.

The gap is the word “reconstruction”, which REA uses and no statute mentions. The established defence for copying functionality is clean-room design: one team studies the original, a separate team writes new code from a specification. An agent that reads the decompiled code and writes the replacement in the same session does both jobs, which critics on Hacker News called washing intellectual property. GeekPark reports that a project rebuilding Photoshop banned REA and Ghidra from its own workflow for exactly that reason. Rights holders are already testing the line. Activision and Bethesda moved against AI-built browser versions of their games in early October.

There is a quieter problem for anyone using REA at work. The binary stays local, but decompiled code, strings and call graphs go to the model provider. If the software belongs to a client or arrived under a non-disclosure agreement, sending it to an American AI company may breach that agreement before any question of copyright arises.

What software companies should protect now

Assume that anything you ship to a customer’s machine can be read. That covers desktop apps, mobile apps, the JavaScript in your web front end, and firmware in devices you sell. Pricing rules, licence checks, scoring models and API keys that live in that code are now visible to anyone with a coding agent and an afternoon. Move the logic that matters to your servers, where an agent cannot reach it. The Korean bank breaches this month went through peripheral systems that were assumed to be out of sight.

Check your licence terms against EU law. A blanket ban on reverse engineering is partly unenforceable in the EU, so rely on what the directive lets you restrict: the use of what was learned and the building of a substantially similar product.

Run REA, or one of the older connectors, against your own products before someone else does. If an agent can explain your differentiating feature in an hour, a competitor can too. Your data and your customers’ trust are harder to decompile.

If your developers want to use these tools, write a short policy first. It should say which software they may analyse, whether client code may be sent to a model provider at all, and that anything rebuilt from another company’s program goes past a lawyer before it ships.