Why Modern Creator?

Mark Kashef · YouTube

Uncensored AI + REA Can Reverse Engineer Anything

An uncensored local model paired with a Ghidra-powered MCP server decompiles QuickTime Player and rebuilds it with the playback controls the frontier models wouldn't touch.

Posted

October 12th

yesterday

Duration

06:38

Format

Demo

educational

Views

17.4K

327 likes

Big Idea

The argument in one line.

Reverse-engineering a desktop app into a working clone no longer requires reading source code or a frontier model's cooperation, just a free decompiling MCP server and an uncensored model willing to execute every step.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You build with Claude Code, Codex, or similar agents and have hit a refusal when asking one to inspect or recreate an existing app's binary.
You're curious what local or uncensored open-weight models (Qwen, GLM) can do that hosted frontier models won't.
You want a cheap, concrete example of agent-driven software cloning, including what it actually costs in API spend.
You're evaluating whether to run inference locally versus through a pay-as-you-go uncensored-model router.

SKIP IF…

You need production-ready, legally cleared software, not a personal proof-of-concept clone.
You're looking for a no-code tutorial. This still requires comfort installing a harness, an MCP server, and a decompiler.

TL;DR

The full version, fast.

Frontier coding models like Claude and Codex often refuse or hedge on reverse-engineering tasks, so this video swaps in an uncensored model to get full use out of REA (Reverse Engineer Anything), a free MCP server that wraps Ghidra for static binary analysis. The stack is three pieces: DeepSeek Harness (an open-source local agent workbench you can swap models into), an uncensored brain (a local Qwen/GLM uncensored model, or a hosted one via Abliteration AI or Orca Router), and REA itself, installed with one prompt. Fed QuickTime Player as a target, the agent decompiled it with Ghidra and rebuilt a clone called Tempo with granular playback-speed controls QuickTime lacks, roughly 80% complete in one shot for about $4, with another $5-6 to polish the look and feel. The ceiling is real: anything that depends on a remote server, like PowerPoint's SmartArt library, can't be replicated locally.

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.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:38

01 · AI app clones and the REA wave

Open-source clones of Photoshop, Word, PowerPoint, and Excel have surfaced via REA, but frontier models like Claude and Codex come with guardrails that limit what the tool can do.

00:38 – 01:45

02 · Rebuilding QuickTime with the controls I wanted

Shows the before/after of a rebuilt QuickTime Player. The original only offers five fixed playback speeds; the rebuild (GLM-5.2, mostly one-shot) adds a renamed app with a new icon and finer speed control.

01:45 – 02:16

03 · How REA works

REA is an MCP server built on the free tool Ghidra that lets an AI agent decompile an installed app's packed binary down to its buttons, functionality, and end state.

02:16 – 03:17

04 · What software clones still miss

Clones get roughly 80% of the way there but can't replicate features that live on a remote server, like PowerPoint's SmartArt diagram library; buttons and local functionality clone fine, server-backed assets don't.

03:17 – 04:22

05 · The three-part uncensored AI setup

The stack is three swappable pieces: DeepSeek Harness (the workbench/limbs), an uncensored model (the brain, local via Qwen/GLM Uncensored or hosted via Abliteration AI/Orca Router), and REA (the tool).

04:22 – 05:00

06 · Set up DeepSeek Harness and your model

Feed the DeepSeek Harness repo to a coding agent to spin it up on localhost, then add a model to the dropdown via a hosted uncensored-model API key from a service like Abliteration AI or Orca Router.

05:00 – 05:28

07 · Install REA and choose an app

A single prompt with the REA GitHub link installs it in under a minute; the agent is then given a target app (QuickTime Player) and gets to work locally.

05:28 – 05:53

08 · Testing my custom player, Tempo

The finished clone, renamed Tempo, opens like any native app. About 80% was one-shot; remaining work was font/look-and-feel tweaks and adding export-format options and granular speed/pitch control.

05:53 – 06:26

09 · What the build cost

The first functional version cost about $4 in model spend; refining the look and feel to the final version cost an additional $5-6.

06:26 – 06:38

10 · Get the full setup guide

Points viewers to a free written setup guide and closes with a like/comment ask.

Atomic Insights

Lines worth screenshotting.

Frontier models like Claude, Codex, and Gemini will work with reverse-engineering tools like REA, but they add friction and guardrails an uncensored model doesn't.
REA is an MCP server that wraps Ghidra, a free decompiler, so an AI agent can peel back a packed app binary down to its buttons, functionality, and end state.
The uncensored-model setup has three swappable pieces: DeepSeek Harness as the workbench, an uncensored model as the brain, and REA as the tool the brain uses.
A GLM-5.2 uncensored model one-shotted roughly 80% of a QuickTime Player clone, including a new name and icon, and installed it automatically.
App clones built this way hit a hard ceiling wherever the original feature lives on a remote server rather than in the local binary, like PowerPoint's SmartArt diagram library.
Nothing in the REA approach is technically new, it just centralizes and packages tools an agent would otherwise have to be told how to assemble itself.
If a local machine can't run an uncensored model well, pay-as-you-go routers like Abliteration AI or Orca Router give API access to the same uncensored models without self-hosting.
The first working version of the cloned app cost about $4 in model spend, and going from functional to polished cost another $5 to $6.
Installing REA into an agent only takes one line, after which the agent can target any locally installed application for analysis.

Takeaway

Why an uncensored model unlocks reverse-engineering tools

WHAT TO LEARN

Guardrails on frontier models like Claude and Codex create real friction on reverse-engineering tasks, and swapping in an uncensored model inside a local agent harness removes that friction for a few dollars a run.

REA wraps the free decompiler Ghidra in an MCP server, so an agent can inspect a packed app's buttons, functionality, and end state without source code.
A modular stack beats a single smart model: an open-source harness (the workbench), an uncensored model (the brain), and a purpose-built tool (REA) can be swapped independently.
Expect roughly 80% completion on a first pass. Anything that depends on a remote server inside the original app, like a cloud-hosted asset library, won't come across in a local clone.
If local hardware can't run an uncensored model well, pay-as-you-go routers like Abliteration AI or Orca Router give the same uncensored access without self-hosting.
Cost is a real data point here: roughly $4 got a first working clone, and $5-6 more got it polished, which is a useful benchmark for any similar agent-driven rebuild task.
None of the individual pieces (decompilers, uncensored models, agent harnesses) are new. The value is in centralizing them so the agent can focus on executing instead of figuring out what tools it needs.

Glossary

Terms worth knowing.

REA (Reverse Engineer Anything): An open-source MCP server that gives an AI agent tools to decompile and inspect installed desktop applications, built on top of the free decompiler Ghidra.
Ghidra: A free, open-source software reverse-engineering tool originally developed by the NSA, used to disassemble and analyze compiled binaries.
Abliteration: A technique for stripping the built-in refusal behavior out of an open-weight language model, producing a version that will attempt tasks a stock model would decline.
DeepSeek Harness: An open-source local agent workbench that runs on a user's own machine and lets any model, local or API-based, be plugged in and given tools.
Orca Router / Abliteration AI: Pay-as-you-go API services, similar to OpenRouter, that give hosted access to uncensored open-weight models without needing to run them locally.

Resources

Things they pointed at.

00:00toolREA (Reverse Engineer Anything) ↗

03:05toolGhidra

03:35toolDeepSeek Harness

03:59toolHugging Face

04:06toolAbliteration AI

04:11toolOrca Router

06:26linkFree REA Setup Guide ↗

06:26productBuild, Learn, Earn with AI (Skool community) ↗

Quotables

Lines you could clip.

05:29

“If the uncensored model is the brain in the jar, then everything else in the harness are all the limbs that the brain can use to inspect and use something like REA.”

crisp, visual metaphor that explains the whole architecture in one line→ IG reel cold open↗ Tweet quote

02:16

“It is pretty efficient, but it's not 100%, and many times it will create a clone that is devoid of some missing features.”

honest limitation statement, builds credibility→ newsletter pull-quote↗ Tweet quote

05:28

“For the first version, it only took $4.”

concrete, surprising cost number→ TikTok hook↗ Tweet quote

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.

metaphoranalogy

So in the past week alone, we've had open source versions of Photoshop, Premiere Pro, Microsoft Word, and even things like PowerPoint and Excel. It seems that every time I open X, there's yet another app being reverse engineered with this new repo called Reverse Engineer Anything. Now you might have seen it used with things like Claude and Codex, but working with those language models comes with a catch.

Sometimes they're not willing to actually execute everything that RIA has to offer. So in this video, I'm going to show you how you can use an uncensored or jailbroken model to get the most out of RIA and not deal with any of the guardrails or opposition that you might get with Frontier models. If you don't believe me, then let me show you.

All right, so what you're looking at on screen is the before and after of rebuilding something called QuickTime Player. It's been a staple app to help you replay any video, whether it's an MP3, .mov, or what have you. Now, I use this app to re -watch the final version of any YouTube video before it goes live and you see it.

But it has a flaw. And the flaw is, let's say I want to vary the speed at which I'm playing it back. I only have a few options I can pick from.

Double, faster, fast, normal, or half. Now obviously there are other softwares out there that I can use that have this feature. But let's say I really like the UI and the simplicity and the elegance of this app.

And all I want is a carbon copy that has this additional feature. So because this GitHub repo is so exhaustive in its approach, you actually don't need a smart model to orchestrate and reverse engineer whatever application you're looking for. And you'll see at the very bottom of my screen that GLM 5 .2 was essentially able to one -shot the majority of this application, come up with a brand new name, a brand new image for the app, and install it on my computer.

Now I'm going to walk you through the simple stack that you need to replicate this, but first, just in case you don't know how it works, let me give you a quick TLDR primer on how RIA works. So let's say you have an app on your computer that you want to replicate. You can use it, but you don't have the source code at your fingertips to easily configure your own version.

What you do have is the finished product, the packed up version that only the computer reads. So RIA essentially allows your AI agent to peel back the curtain and take a look at what's inside that app. It's an MCP server that works with a free tool that's called Ghidra.

And this is the tool that acts like Sherlock Holmes and starts to deconstruct every part of the application from the buttons to the functionality to the end state and how everything comes together. Now, unlike what the average creator is advertising, it is pretty efficient, but it's not 100%, and many times it will create a clone that is devoid of some missing features.

A really good example is if you take the PowerPoint clone that was released this week, and you go to something like SmartArt, which is PowerPoint's way of allowing you to create a diagram or a flowchart. That diagram or flowchart library lives on a server, so you can't replicate that part. You can replicate the buttons, the ability to create shapes, and all the other things like adding text and images.

So in many cases, it'll take you 80 % of the way there, but not necessarily that last mile. Now, realistically, there's nothing in this repo that is revolutionary. It's just a matter of packaging and centralizing all the tools that a language model could use to focus on executing versus focusing on how to reverse engineer.

And again, you can use this with Claude and Codex and Gemini, but you will find some friction in trying to reverse engineer certain types of apps. And I would imagine in the not so distant future, they'll add some patches in their guardrails to stop you from being able to use these open source repos. And like I said before, the setup is unbelievably simple.

We have three core pieces that snap together to make this possible. The first piece is the deep seek harness. It's free and open source, and it's basically the workbench that your AI agent will live in.

So if the uncensored model is the brain in the jar, then everything else in the harness are all the limbs that the brain can use to inspect and use something like Rio. And like you'll see in a second, the beauty of the DeepSeek harness is it lets you swap in whatever model you want. And then you can add your own tools on top, very similar to Lego blocks.

The second piece I alluded to is the brain. And you can use a local uncensored model on your computer. You can go to something like Hugging Face and find something like Quen Uncensored or GLM Uncensored, assuming you can run it.

But assuming that your computer doesn't have the juice to run these models at a high level, you can use something like Obliteration AI or Orca Router. Both of these are basically like Open Router, but they give you access to these uncensored models without you having to host it yourself. And last but not least, the final piece is RIA, and all you have to do, like you'll see in one second, is literally give one line to your agent of choice, and it will install it and plug and play.

So step one is to take this deep seek harness repo, feed it to your language model of choice, and ask it to spin up the harness on local host, basically locally on your computer. Once you run that, you'll see something like this. And if you want to add language models, all you have to do is tell it, hey, I want to run this model on my computer and I want to make it available through this dropdown.

And like I did, you could go to the website's obliteration AI or Orca router. I'm not affiliated with either of them.

And then both will allow you to do a pay as you go plan, like $5, $10, exactly like open router. And then you can just feed the API key tell to integrate it into your deep sea harness. And then it's ready to go.

Once you're all set up, you literally just have to send a prompt like this, which is install this and help me use it to reverse engineer an app. And then I just simply give it the get help link. And even if you're using something like GLM 5 .2 in the cloud, it will install it in under a minute, and all you have to do is tell it the target.

And once I told it, my target was Quick Player. Since this harness is running locally on your computer, it has access to all of these applications and can get to work right away. And depending on what model you're using, within 20 to 30 minutes, you'll have the first version ready to go.

And just like any other app, you'll be able to right -click, click on Open With, I'll click on Tempo here, and 80 % of this was in one shot. The only things that I added on after with a couple extra prompts was basically changing the font, the look and feel, and allowing me to export to different formats. And from the get -go, I was able to increase the speed to whatever I wanted at a much more granular level, see the audio waves, change the pitch, and do anything that I could do in QuickTime Player.

And if you want a sense for the cost, for the first version, it only took $4. And to go from the first version to the final version was an additional $5 to $6. But that was me being very meticulous about the look and feel, but it was functional by the first version.

So although I was initially on the fence as to whether or not this repo was actually helpful, after tinkering around, it is reasonably helpful at centralizing all of these tools. Again, it's not brand new, it's not revolutionary, but it's a convenient way for your agents to do less work to figure out what needs to be done and what tools are required and just focus on executing.

Now, in case you want my exact setup, I'll make that available to you in a full guide in the second link in the description below. And outside of that, as always, if you found it helpful, I'd super appreciate a like on the video and a comment down below, and I'll see you in the next one.

The Hook

The bait, then the rug-pull.

In the past week alone, open-source clones of Photoshop, Premiere Pro, Word, PowerPoint, and Excel have all surfaced, built with a GitHub repo called Reverse Engineer Anything. The catch: frontier models like Claude and Codex won't always cooperate with it, so this video walks through swapping in an uncensored model to get the full tool working.

Frameworks

Named ideas worth stealing.

03:17list

The three-part uncensored AI stack

DeepSeek Harness (the workbench/limbs)
An uncensored model (the brain)
REA (the reverse-engineering tool)

A modular setup where an open-source local agent harness hosts a swappable uncensored model, which then gets REA installed as a tool to decompile and rebuild target applications.

Steal forany workflow that needs an agent to execute tasks a guardrailed frontier model would refuse or hedge on

CTA Breakdown

How they asked for the click.

VERBAL ASK

06:26link

“In case you want my exact setup, I'll make that available to you in a full guide in the second link in the description below.”

Soft CTA to a free lead-magnet guide (Gumroad), paired with a like/comment ask; the paid community (Skool) and consulting offer (promptadvisers.com) sit only in the description, not spoken.

MENTIONED ON CAMERA

00:00toolREA (Reverse Engineer Anything) ↗

06:26linkFree REA Setup Guide ↗

06:26productBuild, Learn, Earn with AI (Skool community) ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.

OTHER LINKSAlso linked in the description.

Storyboard

Visual structure at a glance.

open

hookopen00:00

demo

promisedemo00:38

limitation

valuelimitation02:16

setup

valuesetup04:22

payoff

valuepayoff05:28

CTA

ctaCTA06:26

Frame Gallery

Visual moments.

open

Frame at 00:07 from Uncensored AI + REA Can Reverse Engineer Anything

Frame at 00:12 from Uncensored AI + REA Can Reverse Engineer Anything

Frame at 00:17 from Uncensored AI + REA Can Reverse Engineer Anything

Frame at 00:22 from Uncensored AI + REA Can Reverse Engineer Anything

Frame at 00:27 from Uncensored AI + REA Can Reverse Engineer Anything

One-click upgrade to your Google

Get more breakdowns in your search results

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.

Watch next

More from this channel + related breakdowns.

17:09

Mark Kashef · Review

I Tested Every GPT-6 Astra Effort Level. Here's What I'd Use

Seven parallel Codex runs, one prompt, six effort levels plus a Sol control: does turning up reasoning effort actually make GPT-6 Astra better, or just slower and more expensive?

September 8th

Video of the Day08:38

Mark Kashef · Tutorial

I'm Done Paying Twice for AI

A walkthrough of swapping a metered API key for a headless Codex/Claude subscription so internal tools stop billing per click.

August 19th

08:16

Mark Kashef · Tutorial

Stop Guessing Which Model to Use. Build THIS Instead.

A Claude Code creator mines his own chat history into a personal test pack, then builds a slash-command benchmark that tells him in one run whether a new model release is actually worth switching to.

July 26th

Video of the Day14:13

Mark Kashef · Tutorial

Don't Use Claude Fable 5 Until You See This

A 14-minute cost-routing playbook for the most powerful and expensive model Anthropic has ever shipped.

June 11th

24:39

Mark Kashef · Tutorial

Every Claude Code Secret Its Creator Just Revealed

A 24-minute practical walkthrough of the 15 features Boris Cherny (Claude Code creator) flagged in his 2M-view tips thread.

March 31st

12:22

Mark Kashef · Tutorial

You've Never Used Claude Code Skills Like This

A 12-minute live demo of a Claude Code mega-skill that routes frontend, copy, and bug-fix tasks to Gemini, Codex, or any OpenRouter model while Claude stays in the driver seat.

March 6th