Reverse Engineer Anything: Testing the Viral AI Decompiler Everyone's Hyping
Nick Saraev races a vanilla Claude agent against REA, the trending GitHub tool for AI-powered app cloning. He finds the underlying tech is decades old, and the agents themselves are already commoditizing the software built around them.
Posted
3 days ago
Duration
Format
Review
educational
Views
191.3K
4.4K likes
57 · 43
Big Idea
The argument in one line.
Reverse Engineer Anything (REA) wraps three decades-old decompilers around an AI agent, but a vanilla coding agent can already clone most apps without it, meaning the real shift is software value moving up a layer into the agents themselves.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You build AI agent tools or wrapper products on top of Claude or Codex and want to know if your layer is at risk of getting absorbed.
You're curious whether Reverse Engineer Anything is worth learning or just a trending GitHub repo riding a hype wave.
You want a realistic, tested read on what AI-assisted reverse engineering can and can't do right now.
SKIP IF…
You're looking for a hands-on REA installation tutorial. This video is a skeptical review, not a how-to.
You want proof that reverse-engineered clones are safe to sell. The video explicitly flags real intellectual-property exposure.
TL;DR
The full version, fast.
REA (Reverse Engineer Anything) is an MCP that lets a coding agent decompile any app's binary, using Ghidra, Hopper, and IDA, then rebuild its features from the extracted logic. Saraev tested it by racing REA against a vanilla Claude Opus agent, both tasked with cloning Mac's Preview app from scratch. The REA version matched Preview closely but ran slower and took longer; the unassisted agent finished first with a near-equivalent result. His conclusion: none of REA's component tools are new, AI agents can already reverse-engineer apps through their own reasoning and screenshots, and the real story is software value shifting from individual apps up into the general-purpose agents that can rebuild them on demand.
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.
Cold open on the online reaction to REA: tweets calling it the end of proprietary software, a Replit CEO prediction that nearly everything goes open source, and Saraev previewing his skeptical take.
01:08 – 02:44
02 · How REA Works
Breaks down the actual mechanism: an agent sends a request to the REA MCP, which hands it to a toolkit of decompilers (Ghidra, Hopper, IDA) plus Chrome DevTools, producing pseudocode and proof that the agent loops on until the app is rebuilt.
02:44 – 04:46
03 · Rebuilding Preview
Saraev runs two Claude sessions in parallel inside a terminal multiplexer (Orca), one using REA to clone Mac's Preview photo viewer, one cloning it the vanilla way with no decompiling library.
04:46 – 05:26
04 · Matching the App
Compares the two finished clones against the real Preview app feature by feature, title bar, zoom, share, drawing tools, rotate, save, noting the REA build nailed the features but its drawing tool lagged behind the vanilla agent's.
05:26 – 07:27
05 · Not Exactly New
Points out none of REA's underlying tools are new: Ghidra shipped from the NSA in 2019, Hopper dates to 2011, IDA to the 1990s, and argues the viral GitHub-stars growth chart every REA video leads with is a tell that the hype outpaces the demos.
07:27 – 10:46
06 · What It Means Next
Zooms out to the bigger trend: software is becoming commoditized because value has moved up a layer into general-purpose agents, wrapper apps around Claude and Codex are getting swallowed by base-model updates, and most of what REA does, a vanilla agent will be able to do on its own within a month or two.
Atomic Insights
Lines worth screenshotting.
Reverse Engineer Anything (REA) is an MCP that lets an AI agent decompile any app's binary and rebuild its features from the extracted logic.
REA wraps three decades-old reverse-engineering tools, Ghidra (2019, built by the NSA), Hopper (2011), and IDA (from the 1990s), rather than containing any new technology of its own.
You don't need REA at all: a coding agent can already reverse-engineer a UI by taking its own screenshots and reasoning about what it sees.
In a head-to-head test, Claude Opus rebuilt a clone of Mac's Preview app without REA and finished before the REA-assisted version did.
The REA-built clone matched Preview's native features (zoom, draw, rotate, save) almost exactly, but its drawing tool ran slower than the vanilla-agent version.
Most videos hyping REA lead with a GitHub-stars growth chart instead of a working demo, a sign the story is more cultural moment than technical breakthrough.
The real shift is abstraction, not reverse engineering: software value has moved up a layer, from specific apps to the general-purpose agents (Claude, Codex) that can rebuild those apps on demand.
That abstraction shift means most front-end and back-end wrapper products built around Claude or Codex are being swallowed by the base models' own feature updates.
Reverse-engineered clones carry real intellectual-property exposure, so the practical upside is closer to recombining ideas faster than building a sellable competitor.
Expect agent-native reverse engineering to arrive as a built-in model capability within a month or two, driven by ongoing post-training rather than bolted-on tool libraries like REA.
Takeaway
Agents can reverse-engineer apps without any special tool.
AGENT COMMODITIZATION
REA's real value is proof that general AI agents are already absorbing the specialist tools and wrapper products built around them, not a new reverse-engineering breakthrough.
01REA Is Trending
REA is an MCP that connects an AI coding agent to decompiling tools so it can rebuild a piece of software by studying its compiled binary.
Prominent AI commentators, including a startup CEO, reacted to REA by predicting most proprietary software will soon become open source by default.
02How REA Works
The mechanism is simple: your agent asks REA how a feature works, REA routes that request to a toolkit of decompilers, and returns pseudocode plus proof.
The agent then loops on that output, repeatedly querying the toolkit, until it has reconstructed the whole app from reverse-engineered pieces.
03Rebuilding Preview
Running two coding agents side by side, one with REA and one without, is a fast way to test whether a tool is actually pulling its weight.
The unassisted agent, using only its own reasoning and screenshots, finished cloning a real desktop app before the REA-assisted version did.
04Matching the App
A reverse-engineered clone can match a target app's visible features closely, title bar, zoom, share, draw, rotate, save, while still running measurably slower under the hood.
Feature parity alone doesn't prove a tool's necessity; the vanilla agent produced equivalent functionality without ever touching a decompiler.
05Not Exactly New
Every tool inside REA predates the current AI moment by years: Ghidra (NSA, 2019), Hopper (2011), and IDA (1990s) are all established, mature software.
A video that leads with a GitHub-stars growth chart instead of a working demo is usually signaling that the hype has outrun the substance.
06What It Means Next
The real shift is abstraction: value has moved from individual software applications up into the general-purpose agents that can rebuild them on demand.
Wrapper products built around Claude or Codex are getting absorbed into those models' own feature updates, which shrinks the market for thin AI add-ons.
Reverse-engineered clones carry real intellectual-property exposure, so the practical opportunity is recombining ideas faster, not selling a cloned competitor.
Expect most of what REA does to be absorbed into vanilla coding agents within a month or two as ongoing post-training closes the gap.
Glossary
Terms worth knowing.
REA (Reverse Engineer Anything)
An MCP tool that connects an AI coding agent to decompilers so it can rebuild an app's features from its compiled binary.
MCP (Model Context Protocol)
A standard interface that lets an AI agent call external tools, like REA's decompiler suite, as part of its workflow.
Ghidra
A free reverse-engineering and decompiling tool originally released by the NSA in 2019, now one of the engines REA calls on.
Hopper
A paid macOS disassembler released in 2011, used to inspect a compiled binary's instructions line by line.
IDA
A reverse-engineering and disassembly tool dating back to the 1990s, one of the oldest tools in REA's stack.
Decompiling
Converting a compiled app's machine code back into readable pseudocode so a human or agent can study how it works.
Pseudocode
A high-level, plain-English description of what a piece of code does, without being runnable code itself.
Resources
Things they pointed at.
01:41toolGhidra
01:41toolHopper
01:41toolIDA
02:29toolOrca (terminal manager)
00:00toolClaude Code / Claude Opus
02:36toolCodex
Quotables
Lines you could clip.
00:27
“In this video, I'm going to run you guys through how REA works, but also why I think a couple of people are missing the point.”
frames the whole video's skeptical angle in one line→ TikTok hook↗ Tweet quote
07:40
“You know that when a video starts talking about how many GitHub stars a thing has or how super trending it is, that it probably lacks something more interesting to show you instead.”
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
metaphor
Reverse Engineer Everything, or REA, is an MCP that allows you to virtually rebuild any software you want by actually decompiling the source app and then giving your AI agent a bunch of instructions behind how to recreate and replicate those features. Pretty much everybody is talking about it and doing cool demos like reverse engineering their Apple TV remote in around 45 minutes.
Or, like one person on X thinks, reproduce the entire suite of Photoshop apps. In this video, I'm going to run you guys through how REA works, but also why I think a couple of people are missing the point, and ultimately speaking, what this means for software more broadly. So yeah, everybody on planet Earth is making a big fuss right now about REA.
Matthew Berman said reverse engineer anything, software is done. You know, this fella here said software is dead, long live software. MJAD, the Replit CEO, said this is insane, pretty much all software will soon be open source.
And Jaden Williams says God help us all. You know, as somebody that used to develop software, I can empathize with a lot of these concerns. But there are a couple of things that I think you need to know about REA to really be able to deduce whether or not that is A, true, and B, what you can do about it.
So first, the way reverse engineer anything actually works is you, a lowly human, will ask your super smart squiggly AI agent, hey, how does this feature work? Can you make me this app and reverse engineer it? Your AI agent then sends a request to the REA MCP.
This then sends a request to a tool suite containing apps like Gidra, which I think was initially released in 2019, Hopper, which I believe is 2011, and even IDA, which I'm pretty sure was released in the 90s. This REA library then allows your agent to seamlessly communicate with each of these tools. It can then use Chrome dev tools and other services to help poke around the app and learn more about it before finally giving you some sort of result.
That result usually includes some pseudo code, which just means not actually code, just high level descriptions of what's going on, some proof of what's happening and then any unknowns. Your agent then picks that up and then it just continues circling around and around like a fat merry -go -round until your agent has effectively rebuilt the whole app.
It's pretty easy to get up and running with this stuff. You literally just tell your agent, hey, go download reverse engineer anything, REA, and let me know when you're done. I'm using this inside of a cool terminal manager called Orca.
The reason I like Orca is because I just get to, you know, really quickly and easily create a bunch of terminal subdivisions, and then I can run clod in one, codex in another, and so on and so forth. You don't need to do that. Obviously, you can do this in whatever freaking app you guys want.
I just, you know, I like this one because it makes me feel all smart and big -brained. As you guys can see, the main agent here has already found REA, and now we can go and we can try rebuilding something. All right, now that it's all downloaded and installed, thanks, Claude, we can actually just start a new Claude session.
So I'm going to exit out of this. I'm going to go Claude again. We were indeed bussing.
And I'm going to say use REA to rebuild the preview app on my Mac. So for those of you guys that don't know, preview is just like a little photo app. It looks just like this.
And in my case, I have this loaded up with a pretty badass looking picture of some scary Titan looking thing attacking some tiny human looking thing. And, you know, there's a bunch of different functionality. I have a little eye button that I can click.
I have a little zoom out, zoom in. This is an application, essentially. And so I could, you know, rebuild my own, obviously.
But what if I really liked the way that this was laid out, the way that it looked? What if I wanted the ability to, I don't know, draw boxes or maybe do a little bit of handwriting on it? Well, that's where something like ARIA would come in handy.
And so I just said, rebuild the preview app on my Mac. And now I'm just going to sit around and wait and let's see what happens. One tiny tweak.
While this is happening, I'm going to get Claude to try and rebuild the same app without using REA. So I'm going to say without using REA, rebuild the Mac preview app, the photo viewer one, and then serve me the result. You can do this in, I don't know, however you see fit.
Let's see which one here I like more. It's done some decompiling with Ghidra, and I'd say we're probably halfway done. And it looks like these two are actually in a race to the finish, trying to create the preview app in their own independent ways.
Okay, and so the end result was first completed by Opus without any library. As you guys could see here, we have the preview app. It does seem a little bit laggy.
I think that may be because I'm currently developing the other app in the background, but hey. We'll compare them more broadly in a moment. So it's been another, you know, three, four minutes since still waiting for the one on the left, aka the REA version to work.
It's testing by opening and closing a bunch of tabs, flashing alpha on my screen. I think it's actually going to launch now with three files and show me the difference. Cool.
And so here is the app that REA developed. Okay, and then here is my preview app on my computer. And as you guys could see here, it is virtually the same thing.
If I just were to pull this open to approximately the same size as my other computer, as my other window, let's see. We have what looks to be the title. This one is far longer, which I think explains a big chunk of the difference.
We do have the information. We can zoom in, zoom out. We can share this.
It looks like we can also draw on the page. And the drawing is occurring far faster than the other app. It looks like we can also rotate as much as we want, stop the annotation, and then ultimately save this file to my computer, which is neat.
The issue is everybody's treating this like it's a super new thing. But I think like a lot of technological developments, to be clear, it is not. I mean, Gidra over here has been around since, as mentioned, 2019.
It was actually released by the frickin National Security Agency. I don't know if you guys could tell, but that is an American ass looking eagle. The Hopper disassembler, I don't know if it's clear, was also not made any time recently.
That website looks like it crawled out of the Stone Ages. Finally, IDA, although a little more modern, was also released forever ago. And so the point I'm trying to make here is nothing is super new or special.
All REA does is it just basically gives your AI agent access to these tools and a really simple interface. And that is ultimately what allows it to do this reverse engineering. But I want to make things really clear here.
I've seen a lot of people talk about how dope this thing is and how incredible this enables scientific and software advances to be. But I haven't actually seen very good demos. And so I think this is one of those situations where the internet has probably gotten a little bit ahead of itself.
You know, there's a lot of incentives on signal boosting cool developments like this. Obviously, a lot of people have also been doing very dope things with games like, you know, Tony Hawk's Pro Skater in Halo 3 meets Modern Warfare times RuneScape. You know, I myself, I had a reverse engineer moment a couple days ago where I combined themes from Minecraft and then Attack on Titan and made a really dope game.
And, you know, I think it's more of like a memetic cultural moment than it is, genuinely speaking, a new breakthrough. And part of the reason that I think you know this is because virtually every video that talks about REA starts with a graph that looks something like this, which is just how many GitHub stars it's gotten over the course of the last few days.
You can see that it was, you know, probably reasonably steady about a week ago, a week and a half ago. But last few days in particular had been pretty nuts. You know that when a video starts talking about how many GitHub stars a thing has or how super trending it is, that it probably lacks something more interesting to show you instead.
Like if the only thing you could talk about are GitHub stars, then it's like, well, why don't you just show me the super dope demo instead, right? Like, why are we talking about people's perception to a thing as opposed to the actual value of the thing itself? Anyway, that's just me being an old man.
More broadly, you know, there are some really cool reverse engineering applications. And I think as somebody that's been doing a fair amount of this stuff myself over the course of the last little while, I should talk in generalities about what this means for software. Most software is going to become open source.
And that's because what we've done is we've moved up an abstraction layer from specific software applications and how they're done to a general intelligence, a .k .a. your clods, your codexes that can just rebuild software on the fly. To that end, most of the front ends and back ends are also going to be highly commoditized, if not already.
I mean, we're seeing a bunch of really cool things happen, this Cambrian explosion of opportunity and availability. But, you know, I always like to think about things from a business perspective as a business owner and somebody that's made a fair amount of money with this stuff. what new opportunity does this really open versus what doors does this close and this closes a lot of doors not reverse engineer anything to be clear but just the trend of this technology more broadly over the course of the last few years up until now virtually all of the wrapper apps that wrap around clod and codecs and deliver some sort of additional functionality like sweet visuals and hell even my sexy explainer app you guys could see right over here all this stuff is just swallowed up by the models themselves through design updates you know interactive explainer updates um the cool software games and stuff like that that you guys are developing, you know, the actual value ladder has moved.
And so while I think you could probably get some short -term benefits out of reverse engineering your favorite app and trying to sell it on Black Hat World or whatever, obviously you'd face severe intellectual property restrictions. And so it's less we can make a dope business out of this and more the barrier for recombining pre -existing ideas and turning it into something kind of cool has gone down.
And the barrier to like hitting trending on X with a cool demo like this has also gone down. To be clear, you don't need reverse engineer anything to reverse engineer an app. You could actually just tell your agent, hey, I really like XYZ feature in XYZ app.
Could you go and figure out how it works and reverse engineer it for me? And you know what? It'll use its own little screenshot tool, take screenshots of like an animation as it goes across the page or whatever the hell it is that you want it to copy.
It'll do it for you. You don't need a cool library or whatever. It'll already know how to do that just through virtue of its own reasoning process alone.
Will it do so perfectly? Will it? copy the exact functions of the app?
Probably not. But I wouldn't be surprised if a lot of those functions that it is attempting to replicate actually end up being more efficient when written by AI than by humans alone. We saw this with the recent Photoshop suite repurposing, and it's going to be pretty cool.
You know, we can get pretty close with just vanilla cloud code. And I want you guys to know that that's now. I mean, in like a month or two, realistically, cloud code is going to have a lot of these reverse engineering features, not built directly in, but the post -training that is continually occurring on it.
We'll guide it towards being able to do most of this on your own. We also don't know if we need all of the component apps there. Gindra, for instance, or IDA.
Obviously, they did help replicate speed. So I think that's pretty important. We can produce a plausible.
visual and like backend analog of any app with REA. We can also do it visually and like probably two thirds of the time, if not faster, just by using vanilla agents. So while there is a little bit of a speed bump there, I just want you guys to know that most of this would probably be just captured by agents in general.
So get ready for the reverse engineer era. All right, hopefully you guys appreciated the video. I had a lot of fun putting it together and I will catch all y 'all in the next one.
The Hook
The bait, then the rug-pull.
Everyone from Matthew Berman to the Replit CEO is calling REA the end of proprietary software. Nick Saraev raced it against a vanilla Claude agent to see if the viral decompiler tool actually earns the hype.
Frameworks
Named ideas worth stealing.
01:24model
REA pipeline (agent to REA to tool suite to loop)
Agent asks REA how a feature works
REA routes the request to Ghidra, Hopper, IDA, and Chrome DevTools
Tools return pseudocode, proof, and unknowns
Agent loops on the result until the app is rebuilt
The mechanical loop behind REA: an MCP wrapper that gives an agent structured access to three existing decompilers, no new reverse-engineering technique required.
Steal forany 'agent orchestrates specialist tools in a loop' explainer diagram
CTA Breakdown
How they asked for the click.
VERBAL ASK
10:36next-video
“I had a lot of fun putting it together and I will catch all y'all in the next one.”
Soft sign-off only, no explicit subscribe or link ask on camera despite Maker School and Maker Zero links sitting in the description.
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 creator who says he's burned $31,141 on Claude Code turns that spend into sixteen concrete habits, from running evals on prompts to keeping a backup model on standby.
A creator downloads an ad-funded coding platform promising free daily hours on GPT-5.6 Luna and GLM 5.3 Flash, then builds two live demos to see if the free tier actually produces working code.
A five-step AI pipeline — generated painting, AI video, frame extraction, dithering, and free deployment — that turns a couple dollars of image and video credits into an animated, expensive-looking website hero.
Cerebras published exactly how its internal knowledge base works, and it's a plain retrieval pipeline any team can copy — no graph visualizations, no floating brain, just Slack and Wiki and code stitched into one queryable table.
00:14
00:22
00:37
00:44
00:52
00:57
01:08
01:16
01:24
01:32
01:49
01:57
02:08
02:13
02:21
02:29
02:37
02:53
03:01
03:09
03:17
03:25
03:33
03:41
03:48
03:58
04:06
04:15
04:22
04:31
04:38
04:54
05:02
05:10
05:18
05:24
05:32
05:43
05:54
05:59
06:07
06:15
06:27
06:31
06:36
06:49
06:55
07:12
07:20
07:27
07:36
07:44
07:52
08:00
08:08
08:16
08:26
08:32
08:44
08:48
09:05
09:13
09:21
09:29
09:37
09:45
09:53
10:01
10:09
10:17
10:25
10:33
10:41
25:59
09:41
6:23:04
6:02:49
22:52
23:48