REA, short for Reverse Engineer Anything, has become one of the most closely watched repositories on GitHub after a sharp rise in attention over two weeks. As of Oct. 20, the project had about 340,000 stars and 85,000 forks, up from fewer than 13,000 stars on Oct. 8. On Oct. 16, AI search company Perplexity integrated the tool into its AI agent product, Computer.
What REA does
REA is an MIT-licensed open-source project distributed through the npm package rea-agents. It uses MCP, a standard protocol that lets AI agents call external tools, to connect with Claude Code, Codex, Cursor and Gemini CLI.
The setup lets users issue plain-language prompts such as asking how an app’s search function works, requesting supporting evidence, and then asking for a similar implementation in their own project. REA is designed to let an AI agent inspect a target program and explain how it works even when the source code is unavailable.
REA is not a new decompiler. For native binaries, it calls established tools including Hopper, Ghidra and IDA. Ghidra was open-sourced by the U.S. National Security Agency in 2019. REA’s role is to organize the output from those tools into structured results for AI agents, with each finding paired with its source and stated limitations.
According to the project description, REA can analyze Windows, macOS and Linux executables, JavaScript and Electron desktop apps, .NET programs, Android APKs, websites, firmware and Ethereum contract bytecode. The analysis runs locally on the user’s machine, but the results are passed to the AI agent, which means they also go through the model provider.
Perplexity adds REA to Computer
Perplexity CEO Aravind Srinivas announced the integration on X. The company’s official account said reverse-engineering tasks would be free through Oct. 31, with a limit of 10,000 points per user, and showed five apps built with the feature, including a Figma-like design tool.
Official demos include Notion clipboard tracing and a 1990s game formula
The REA website lists several demonstrations. One traces how copied text in the Notion desktop app moves from the web interface to the system clipboard and identifies the check that allows only Notion’s own pages to write to the clipboard.
Another example covers the sound channel calculation in the older game DX-Ball, determining whether the sound should lean left or right when a brick is hit. The project said an AI agent used REA’s analysis to rewrite the 63-byte function in C, matched all 3,205 input comparisons against the original program, and produced recompiled machine code that matched byte for byte.
Star growth drew questions, while npm downloads also jumped
The speed of the GitHub star increase led to questions in the community about whether the numbers had been inflated. One reference point is installation data. Official npm download statistics show rea-agents had 16 downloads on Oct. 2, climbed to 103,109 on Oct. 12, and stayed above 100,000 per day from Oct. 16 through Oct. 18.
Those download figures may include automated systems and mirror sites, but they still indicate that a large number of users were installing the package.
The project’s popularity also brought imitation risks. REA’s README warns users to download only from the official GitHub repository, the rea.tools website and the rea-agents npm package. Repositories with similar names or copied README files are not official releases.
Independent test found accurate hits, but notable omissions
SystemDesign Academy tested REA 6.4.0 on an open-source Electron sample app, which made it possible to compare every answer against the actual codebase.
In that test, all 30 code locations identified by REA were correct. It also found one communication channel that a text search had missed.
But the test also found gaps. The sample app had eight cases where the main process returned messages to windows, and REA recorded none of them. The program opened eight windows, while REA counted only one. Four files had nodeIntegration enabled, a setting that gives web code access to the whole computer, and none of those appeared in REA’s summary.
The tester’s conclusion was that findings reported by REA could be trusted, but anything not mentioned should be treated as unknown rather than assumed not to exist.
In a Hacker News discussion thread, one commenter also asked: 「What is the difference between this and simply asking Claude to install Ghidra and start analyzing on its own?」
Project says it will not help bypass DRM or anti-cheat systems
As the tool lowers the barrier to reverse engineering, copyright and security concerns have moved into focus. REA’s disclaimer says the project is limited to security research, malware analysis, software preservation, compatibility development, and analysis of software that users own or are authorized to inspect.
The project says it does not support and will not assist with bypassing DRM, license checks or anti-cheat systems. It also says it will not help copy proprietary software, recreate protected code or assets, or evade the usage policies of AI model providers.
The disclaimer adds that reverse engineering may be restricted by local law or software license terms, and users are responsible for obtaining authorization. The report also noted that AI-driven attempts to recreate commercial software are not unique to REA, citing an Oct. 7 report by Chain News on ArtCraft, a free Adobe alternative, while noting that the two are separate projects.


