AI Tools9 min read

Reverse Engineer Anything (REA): What It Is and Why It Exploded

Reverse Engineer Anything (REA) went from 41 to 300K+ GitHub stars in 15 days. What it is, what's actually new, how to install it and what it found in my app.

By Moe Lueker·Oct 19, 2026

Reverse Engineer Anything (REA) is a free, open-source tool that lets a coding agent like Claude Code or Codex look inside software it has no source code for and explain how a feature works. It connects the agent to reverse engineering tools such as Ghidra, Hopper, IDA, a browser and runtime capture through one MCP server. You ask in plain English. You get back a reconstruction of the logic, the evidence for each claim and a list of what it couldn't work out. You don't get the original source code.

It is also the fastest-growing repo I've watched on GitHub. On October 2, 2026 it had 41 stars. On October 19 it had 335.7K. I used it on my own Mac app for a video, so this post mixes the public numbers with what I saw first-hand.

REA's GitHub stars over time#

The repo was created on April 14, 2026, and sat almost unnoticed for months. GH Archive records its first star event on August 14 and only 34 stars by the end of September. Then it went vertical.

The curve has three phases you can read straight off the data:

• Ignition, October 3 to 8. 123 stars became 25,576. Developers found it and shared it.
• The explosion, October 9 to 13. It added more than 19,000 stars a day, peaking at 51,973 on October 12 (UTC). It crossed 100,000 on October 11 and 200,000 on October 13.
• The long tail, October 14 onward. Still 12,000 to 35,000 new stars a day, but falling every day. It passed 300,000 on October 17.

GitHub's own star counter agreed. I read 157,134 stars on the evening of October 12, 320,274 at lunchtime on October 18, and 335,538 on the afternoon of October 19. Each of those sits within about 1% of the GH Archive line.

Stars are a vanity metric, so I also checked installs. The setup package, rea-agents, did more than 100,000 npm downloads a day on October 12 and again on October 16, 17 and 18. People weren't just starring it. They were running it.

Why REA got so popular#

The pitch is one sentence#

"See a feature you like. Understand how it works, down to the binary level." That's the line at the top of the README, and it lands with anyone who has ever wondered how an app does something. You don't need to know what a decompiler is to want that.

The demos did the marketing#

YouTube carried it. As of October 19, Low Level's "I Tested The Viral Reverse Engineering Repo" had 2.35 million views, Nick Saraev's test of "the viral GitHub repo that reverse engineers anything" had 1.53 million, Rob Shocks had 491,000 and Eric Tech's "How To Use REA To Clone Any Software In Minutes" had 337,000.

It also rode a bigger wave. Fireship's "Every app on Earth is now open source" passed 6.2 million views, mostly about the free Adobe rebuilds. Those apps were rebuilt from scratch, not reverse engineered with REA, but the two stories blurred together into one idea: closed software isn't closed anymore.

Coding agents were finally ready#

Since early 2025 an agent could call Ghidra through an MCP server, but you had to wire it up by hand and know what to ask. Coding agents are now good enough to plan a multi-step investigation on their own. Models have improved fast at this kind of work: OpenAI reported 88.0% for GPT-6 Astra on SRE-Bench, a binary reverse-engineering benchmark, against 55.9% for GPT-5.6 Sol. I break those numbers down in my GPT-6 Astra benchmark post.

Setup is one line#

npx rea-agents setup registers REA with your agents and installs matching instructions. No plugin to compile, no Python environment. Low friction plus a fun demo is how a repo gets 50,000 stars in a day.

What's genuinely new, and what isn't#

Most of REA's power is old. Decompilers have existed for decades. The NSA released Ghidra as free, open-source software in March 2019. GhidraMCP, which lets an AI call Ghidra's functions, appeared in March 2025.

What REA adds is the orchestration around those tools:

• One MCP for many target types. Native binaries, JavaScript and Electron apps, websites, .NET assemblies, Android APKs, firmware, saved network captures and more, each routed to the right tool.
• Evidence and unknowns as the output format. Every finding carries the evidence behind it, and the agent can record what it could not resolve instead of guessing.
• Agent instructions included. Setup installs a skill that tells your agent how to run an investigation, not just which tools exist.

Here's how it compares with the tools people used before:

Feature lists come from each project's README on GitHub, read October 19, 2026.

The honest caveat: for simple jobs, a plain coding agent can often do the same work. Two of the creators I studied before my video found a regular agent as fast or faster on small targets. That matched my own read. REA earns its place when you need to see what the code really does, with proof.

How to install REA#

You need Node.js 22.19+ (or 24.11+, or 26+) and npm. Then:

npx rea-agents setup

Pick your agents (Claude Code, Codex, Cursor, Gemini CLI and others are supported), review the changes and approve. Setup backs up your config first. Restart your agent afterward.

For native apps you also need a decompiler. Ghidra is free. Hopper is paid with a demo, and setup can install it after you approve. IDA works if you already have it. JavaScript, Electron and .NET analysis need none of them.

Or skip the terminal and ask your agent to do it. Paste the GitHub link and say: "Install REA from this repo for Claude Code on my Mac, and tell me if it needs Ghidra or Hopper." I put that prompt, a free Ghidra setup prompt and 11 more into a free starter guide.

How to use REA well#

These are the habits that made my run useful:

1. Use an empty folder. Put only the app inside a target folder. If the agent can see your source, it will "discover" what it already knows.
2. Ask about one feature. "How does this app turn text into speech?" works. "Rebuild this whole app" doesn't.
3. Ask for three buckets. Tell it to sort every claim into what it can prove, what it infers and what it can't see. This is where REA beats a plain agent.
4. Raise the timeout on big binaries. The README lists REA_GHIDRA_STARTUP_TIMEOUT_MS and REA_HOPPER_STARTUP_TIMEOUT_MS for large files. I wish I had set it before my run.
5. Keep approvals on. REA runs with your permissions and is not a sandbox. A file you inspect can contain text aimed at your agent.
6. Update often. The project ships fixes constantly. Rerun npx rea-agents@latest setup when something breaks.

What REA found in my own app#

I pointed REA at HoverSpeech 0.1.3, a free text-to-speech Mac app I built, in an empty folder with my source code blocked. Ghidra timed out twice, after 330 seconds each, on the app's 16 MB helper file. So the agent fell back to labels, the file list and type names.

It still identified the speech model (Kokoro-82M through the FluidAudio library), the voice it uses ("Heart"), and that the model downloads on request instead of shipping in the app. Its report had 24 proven claims, 7 inferences and 7 things it couldn't see.

Then Claude rebuilt the read-aloud feature as a web page and put it online. The whole run used about 56 million tokens, roughly $28.60 at API prices, or 2% of my weekly Claude plan. The full step-by-step, with screenshots and every prompt, is in how to use REA: I cloned my own app and put it online. You can try the page it built.

Is REA legal?#

The project's own position: REA is for lawful research, education, interoperability, software preservation and analysis of software you own or are authorized to examine. Commercial tools like IDA and Hopper still need proper licenses. Point it at your own software, or software you have permission to inspect, and read the license of anything you analyze. This isn't legal advice.

If what you really want is your own version of a tool, forking an open-source app is often the cleaner path. I did exactly that when I built my own dictation app with Claude Code and Gladia.

REA FAQ#

The takeaway for anyone who ships software: a compiled app is no longer much of a lock. What protects your product now is what runs on your servers, your users and your distribution.

Some links above are affiliate links, so I may earn a commission at no extra cost to you. I only recommend tools I actually use.