The Open Source Tool That Lets Claude Reverse-Engineer Any App
REA turns Claude Code into a reverse engineer: it maps a binary, an Electron app, or a website down to the machine code, then hands back an evidence-backed write-up Claude can build an original version from.
Posted
today
Duration
Format
Tutorial
educational
Views
9K
126 likes
57 · 43
Big Idea
The argument in one line.
REA gives Claude Code a structured, evidence-based process for deconstructing any application down to the machine-code level, so a developer can build a genuinely custom version instead of guessing at how the original works.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You use Claude Code and want to build your own version of an app you already use, not a 1:1 copy of it.
You're curious how an AI tool can reverse-engineer closed-source software without just hallucinating an explanation.
You work with compiled binaries, Electron apps, or websites and want to understand what's happening under the hood before building against them.
SKIP IF…
You're hoping to clone software you don't already own — REA still requires the target app installed on your machine, so it's not a piracy shortcut.
You don't use Claude Code or MCP tools — this workflow is built specifically around how Claude issues questions and compiles its write-up.
TL;DR
The full version, fast.
REA is an open source MCP server that lets Claude Code reverse-engineer any application: binaries, Electron apps, or websites. It works in four steps. You point it at an app and ask a question, it maps the app by routing each part to the right engine (Ghidra for machine code, a JS analyzer for Electron apps, Chrome tooling for websites), Claude investigates by asking small precise questions that REA answers with evidence labeled seen, inferred, or unknown, and finally Claude writes new code from that write-up rather than copying the original's pseudocode. The demo deconstructs the dictation app Wispr Flow: Ghidra pulled 56,803 functions from its native helper in 27 minutes, and a JS analyzer pulled 1,474 pieces from its Electron app in 7 minutes. The real use case isn't cloning software you already own, it's using the map to build your own customized version.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
REA is an MCP tool, not a general prompt: it picks a different engine depending on what it's deconstructing, so a compiled binary, an Electron app, and a website each get analyzed differently.
Every claim REA generates is tagged seen, inferred, or unknown, so the output comes with its own evidence trail instead of presenting guesses as facts.
REA's repo passed 200,000 GitHub stars within about a week, going from flat to vertical on its star-history chart almost overnight.
The actual value of REA isn't cloning an app you already own, since you need the software installed to point REA at it; the value is using the map to build a customized version for your own use case.
For machine code, REA hands the job to Ghidra, the reverse-engineering framework originally built by the NSA, which disassembles compiled binaries into pseudocode a model can reason about.
In the demo, Ghidra needed about 27 minutes to import the dictation app Wispr Flow's native helper and surfaced 56,803 functions.
The Electron 'brain' of the same app took a JS analyzer about 7 minutes and surfaced 1,474 pieces, running in parallel while Ghidra was still importing.
REA's map for Wispr Flow also surfaced 327 web addresses and a server address the app downloads at runtime, information that wouldn't show up just by reading source.
Claude doesn't rely on REA for everything: on the Electron side of the demo, Claude made 5 REA calls but did about 40 of its own searches directly against the code.
The investigation step is a back-and-forth loop: Claude asks small, precise questions and REA answers each with a fact plus a receipt, and where REA can't reach, Claude reads the code itself.
The final step deliberately avoids copying: Claude writes new code from the write-up, not from the original app's pseudocode, so answered questions become the blueprint and unknowns become the developer's own design choices.
Takeaway
A four-step loop for reverse-engineering anything
WHAT TO LEARN
REA turns Claude Code's black-box guessing about how an app works into a four-step loop: ask, map, investigate, and build, with every claim backed by seen, inferred, or unknown evidence.
Step one is just asking: point Claude at an app via REA and state what you want to know, and REA loads its own playbook to pick the right tool for that type of software automatically.
Step two, mapping, routes each part of the app to a specialized engine: Ghidra for compiled machine code, a JS analyzer for Electron apps, and Chrome tooling for websites, each producing pieces, strings, addresses, and call relationships.
Step three, investigation, is a question-and-answer loop: Claude asks small precise questions, REA answers each with a fact and a receipt, and where REA can't reach, Claude reads the code itself and contributes its own searches.
Step four is building: Claude writes genuinely new code from the write-up rather than copying the original's pseudocode, so answered questions become the blueprint and unknowns become your own design choices.
The labeling discipline (seen, inferred, unknown) on every single claim is what keeps REA from just hallucinating a plausible-sounding explanation of code it hasn't actually verified.
Running different analysis engines in parallel matters for speed: in the demo, the JS analyzer finished the Electron app in about 7 minutes while Ghidra was still 27 minutes into importing the native binary.
The tool is most useful not for cloning software you already own, but for using its evidence map as the foundation to build a version customized to your own workflow or business.
Glossary
Terms worth knowing.
REA (Reverse Engineer Anything)
An open source MCP tool for Claude Code that reverse-engineers binaries, applications, and runtime behavior and returns an evidence-labeled write-up of how they work.
MCP (Model Context Protocol)
A protocol that lets an AI coding assistant like Claude Code call out to external tools and skills, giving it capabilities beyond its own training, such as REA's reverse-engineering engines.
Ghidra
An open source software reverse-engineering framework, originally built by the NSA, that disassembles compiled machine code into assembly and readable pseudocode.
Electron app
A desktop application built with web technologies (JavaScript, HTML, CSS) wrapped in a Chromium shell; common for apps like Wispr Flow, Slack, and VS Code.
Pseudocode
A readable, natural-language-like approximation of a program's logic, generated from disassembled machine code so a human or model can understand what it does without the original source.
Seen / inferred / unknown
REA's three-tier confidence label for every claim it makes: seen means it found direct evidence in the code, inferred means it's a reasoned guess, and unknown means it couldn't determine the answer.
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
This free open source GitHub repo just gave Claude the ability to reverse engineer anything, whether that's a desktop app, a website, or anything else in between. It allows Claude to look all the way down to the machine code of a particular application and figure out how it works. So you can not only clone it, but create a version that is specific and customized to you.
This tool works on over 20 different types of software. The repo itself is at over 200 ,000 stars. And today I'm going to explain exactly how it works so you can start using it yourself.
So Ria has been hyped like crazy this last week. And I think the star history kind of shows that as pretty much invisible all the way till a couple of days ago where it's gone straight vertical. And we're now at 206 ,000 stars.
And by the time I finished recording this, I'm sure we'll be like 210. Why the hype? Well, on the surface, it's like, well, it can deconstruct anything.
Therefore, you can clone any sort of software you want, which is useful, except you realize in order to run Raya and point it at some sort of software, you already have to have that software on your computer. So why am I simply cloning something I already own or something that already exists? The real sell here is, hey, I can look at something that does exist already, but I can now deconstruct it in order to make something.
new, create a customized version of that application that fits my life, my use case, my business, whatever. That's the huge sell for this thing. And Raya is super powerful because it acts as this MCP tool that picks specific tools and skills for the specific type of application you're trying to deconstruct.
It's not a one size fit all thing. It's not a one trick pony. And so let's break down how this thing actually works so you can use it yourself.
And for today's video we're going to be using as a demo whisper flow we're going to say we're deconstructing whisper flow and it is through that lens that i'll show you how rea gets something like this done now the first step is you need to install rea and to do that you simply run this command or you can just copy the url paste that in a cloud code and say okay can you go ahead and install this for me and then the first step is you simply ask rea to begin deconstructing some particular application.
Now that prompts can be very simple. It can look something like this. Hey, can you take the Rhea skill and take a look at Whisperflow?
I want you to break it down into its component parts so you understand how it works so we can create our own custom version of it. And once you do that, it's then going to load the Rhea playbook, aka its skills. Then it's going to pick the right tool for the job depending on that particular application.
Remember, there was like 20 different applications it can work with. It's going to create a summary of what's going on with this thing. And then it's going to start claiming a bunch of things about this application based on your prompt.
So I said, hey, deconstruct whisper flow. So it's going to figure out, hey, what happens when the microphone gets turned on? What are the hotkeys?
How does text get sent from one place to another? How does it go into a text box? And for every single question, it's going to give us a claim.
And that claim will either be seen, inferred, or unknown. If the claim can have some sort of evidence to it, it's going to say, hey, here's the evidence in the code.
If it's inferred, it will spell that out. And if it has no idea, it's going to say it's unknown. So it's not just making stuff up.
We can always see it's works cited. So for WhisperFlow, like I said, it's going to take our prompt and essentially break that down to like, all right, hotkeys, recordings, cleanup, typing work. What's the evidence?
It's going to take a look at WhisperFlow and it's going to understand, well, WhisperFlow is actually two parts. We have the Electron app. And then we have the native helper, the thing that actually works with Windows or Mac to record keystrokes and things of that nature.
Now, before we go into the next step, a quick word from today's sponsor, me. So I just released a brand new Cloud Code Masterclass, and it is the number one way to go from zero to AI dev, no matter your technical background or lack thereof. We focus on real use cases, and I assume you know nothing.
So no matter where you're starting from, I can get you spun up on the most powerful AI tools in the game today. So if you want to get started, there is a link to it in the pinned comment.
Hope to see you there. From there, Rhea is then going to map the application. And so to do that, it's going to send each part to the right engine.
When we say the right engine, we're really talking about the right skill or the right tool for that job. So for example, remember how I said WhisperFlow is actually two parts. It has a helper in the Electron app.
Well, the helper... isn't something that Claude can really look at and understand what's happening. This is something where we need to look at the actual machine code.
This is one of the really cool things about this reconstruct everything application. It gives us the ability to use tools that look at like the actual zeros and ones, so to speak, that are happening at the machine level. And that tool it uses in this case is called Gidra.
Now, Gidra was actually created by the NSA. Yes, that NSA. It is open source, and the whole idea is it looks at machine code, takes the machine code, turns that into assembly, and through assembly, pretty much writes pseudocode, which kind of just means it turns it into natural language.
And what that means for you is that using this skill, Claude can now begin to deconstruct things and figure out how certain applications work at a very base level. And so he looks at all this stuff and essentially creates a map. You know, the pieces, the strings, the web address, who calls what, how is this all connected?
And so with WhisperFlow, it figured out, hmm, using Ghidra, which took 27 minutes, it identified over 56 ,000 functions. And the brain, the Electron app, all that AI stuff, it used a JavaScript analyzer. It took it seven minutes, and it found just under 1 ,500 pieces.
And it turned all that information into a map. that it's going to use when Cloud Code asks it questions to, again, kind of put together how everything works. So we ask Claude using Rhea to look at an app, and now it's mapping that thing out.
It's breaking it out into its component parts and calling on specific tools to do that. Next, step three kind of begins the investigation piece. This is a question and answer loop between Claude and Rhea.
claude asks small and precise questions and ria answers each with a fact and a receipt remember earlier it's making all these claims and it's figuring out where are these answers coming from right it's always going to give us receipts of this information so it doesn't just make stuff up now there are instances where claude actually can do this better than ria itself and claude can go ahead and read code when we talked about that electron app remember it was a javascript analyzer well claude can do that on its own and in the end we get a write -up with those seen inferred and unknown tags now with whisperflow remember we had those two parts we had the helper and we had sort of this brain and claude had questions about both of them when we were talking about questions with the helper and we wanted to look at this information from a machine code level you know zeros and ones well we had 13 rea calls and we got all the information we needed interestingly when we asked questions about the electron app
We got five Raya calls, and we did 40 of our own searches. Claude was able to actually go in here and look at things without necessarily using this GitHub repo, using this skill. And so this is kind of the moneymaker.
This is where we've taken all the information that reverse engineer anything has collected, and we have Claude, again, sort of begin to tease everything out so it eventually can come back to you in step four and say, hey, here's how this thing works. We've reverse engineered it. Here's all the questions I think you would have had about how this thing actually works.
Here's all the evidence for where we got this information. Now let's go ahead and write new code for our custom implementation. That four -step process is how this application works.
We ask it questions. It maps everything out. Claude and Ria have it back and forth to figure out, okay, what's really going on here?
And then in the end, we essentially get a playbook. for our new application that's why this thing is popping off and the secret sauce is this whole thing this idea that it has specific tools for all these different types of applications because if you just go at this with cloud code itself it's a little bit harder you're expecting it to do it on its own instead this github repo has spelled it out specifically for a cloud what's the best way to reverse engineer this specific thing whether it's a go binary a rust binary you know sqlite databases on and on and on and on It's like giving Cloud Code the answers to the test.
So the real question you should have after this video isn't, hey, what can I go clone? It's, hey, what can I go customize and create that's unique to me? So that's where I'm going to leave you guys today.
As always, make sure to check out Case AI Plus if you want to get your hands on your Cloud Code Masterclass. And besides that, I'll see you around.
The Hook
The bait, then the rug-pull.
An open source repo called REA just crossed 200,000 GitHub stars by giving Claude Code something it didn't have before: a structured way to take an app apart, down to the machine code, and hand back a map backed by actual evidence instead of guesses.
CTA Breakdown
How they asked for the click.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A 15-minute breakdown of the two-part feature that lets Claude spawn hundreds of isolated sub-agents for complex tasks and why the default single-session approach fails.
Chase walks a fifteen-minute carousel of GitHub repos — caveman, Graphify, Claude-Video, OpenDesign, CodeBurn, Impeccable, design-extract, CareerOps, browser-harness, n8n MCP — that all dropped in the last thirty days.
Claude Code doesn't generate video pixels. It writes a function that returns one frame for any given time, then a headless browser and FFmpeg turn that function into an MP4 for free.
00:00
00:10
00:17
00:24
00:31
00:38
00:45
00:51
00:58
01:05
01:12
01:19
01:26
01:30
01:40
01:47
01:54
02:03
02:07
02:13
02:21
02:28
02:34
02:42
02:48
02:55
03:02
03:09
03:16
03:23
03:30
03:37
03:43
03:53
03:57
04:04
04:11
04:18
04:23
04:31
04:38
04:45
04:52
04:59
05:06
05:13
05:20
05:23
05:33
05:40
05:47
05:54
06:01
06:08
06:14
06:21
06:28
06:35
06:42
06:49
06:56
07:03
07:09
07:16
07:23
07:30
07:37
07:44
07:51
07:57
08:04
08:11
08:18
08:22
08:32
08:39
08:46
08:52
08:59
09:06
15:13
13:57
20:07
16:27
15:08
14:10