This GitHub Repo Lets Claude Clone Any App (REA)

Author: Hyperautomation Labs · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: This GitHub Repo Lets Claude Clone Any App (REA) · Channel: Hyperautomation Labs · 07:33 · Watch on YouTube

Topics: AI agents, Cybersecurity and software security, Claude Code, Software engineering with AI

Summary

The video demonstrates REA, a tool that gives coding agents ways to inspect software, using Obsidian as a test case. Claude used it alongside direct analysis of Obsidian’s bundled code to reproduce the Quick Switcher’s ranking behavior, matching the original across 4,915 queries. A second test mapped what Obsidian’s window can ask the system to do, but Claude Opus 5.5 stopped the session with a safety warning; Opus 4.8 completed the analysis. The presenter stresses that REA provides evidence and maps rather than original source code, notes the substantial disk space its evidence files can require, and cautions viewers to respect software licenses.

Key points

REA is an MCP server offering 139 tools for coding agents such as Claude Code, Code X, and Cursor.
The video tests REA on Obsidian, an Electron app whose code is bundled inside the application.
REA produced a broad analysis of the bundle, but its map did not index the string needed to trace the Quick Switcher, so Claude examined the minified JavaScript directly.
Claude reconstructed the Quick Switcher’s matching and scoring behavior and compared it against the original over 30 note titles and 4,915 queries.
The reproduced page matched the original’s rank lists, scores, and highlights in every tested query.
Opus 5.5’s safety classifier interrupted an attempt to map Obsidian’s system-facing capabilities; Opus 4.8 completed the task.
The resulting map identified 42 message channels, including handlers involving file operations, web requests, and a shell command.
The presenter says REA does not recover original source code; it supplies pseudocode, maps, call paths, and evidence, and its evidence files can be large.
The video distinguishes learning how software works from copying code or breaking license checks, and advises checking the target app’s license.

Chapters

Start	Chapter	Summary
00:00	Introducing REA and the Obsidian test	The presenter introduces REA and explains how it can help coding agents inspect software, then sets up a test using Obsidian.
02:10	Reconstructing Obsidian’s Quick Switcher	After REA’s initial trace comes up empty, Claude reads the bundled code, rebuilds the matching behavior, and validates it against thousands of queries.
04:18	Mapping system capabilities and a safety interruption	Claude Opus 5.5 stops while analyzing Obsidian’s message handlers, while Opus 4.8 completes a map of the app’s capabilities.
05:57	Other REA examples and limitations	The presenter shares REA case studies and explains its limits, the disk space its evidence can take, and the importance of respecting licenses.
07:01	License caution and closing question	The presenter reiterates the distinction between studying software and violating licenses, then invites viewers to consider testing their own apps.

Quotes

4 minutes. That’s how long it took Claude to take apart Obsidian search with no source code and rebuild it so exactly that it matched the original on 4,915 tests. — 00:00

It doesn’t give up, and it doesn’t make something up. — 02:45

Same rank lists, same scores, same highlights every single time. — 03:47

The model itself decided that mapping what an app can do to your machine looked too much like an attack. — 04:51

Every closed app on your computer just became readable. — 06:30

Concepts and entities

REA, Obsidian, MCP, Claude Code, Opus 5.5, Opus 4.8, Ghidra

Full transcript (original language)

00:00 4 minutes. That’s how long it took Claude to take apart Obsidian search with no source code and rebuild it so exactly that it matched the original on 4,915 tests. Then I asked it one more question about the same app and Claude’s own safety system stepped in and stopped it mid-sentence. The tool that did both is called REA, reverse engineer anything. It passed 76,000 stars this week and went to number one on GitHub trending. It plugs

00:32 into Claude code, code X or cursor and lets your coding agent look inside almost any app. This is the whole story. What it found, where it fell flat, the moment it got stopped and what that means for every app on your computer including yours. Here’s the problem every builder has. You use an app every day and it has one feature you’d love in your own project, but it’s a sealed box, no source code. You can see what it does and nothing

01:04 about how. REA’s promise is on the first line of its read me. See a feature you like, understand how it works down to the binary level and build your own version. It’s an MCP server, the standard way to hand a coding agent new tools and it hands yours 139 of them. Electron apps, websites, .net, Android, firmware and native Mac, Windows and Linux programs if you have Hopper or Ghidra. Ghidra is free. Setup is one line. NPX REA agents setup. It shows you a plan, backs up your

01:37 config, you approve, you restart your agent. I put it in a fresh folder on my Mac with Claude code running Opus 5.5 and I picked a target. Obsidian, the note app. It’s an electron app, so it’s real code ships as one 24 megabyte bundle inside the app. I copied that bundle into the project and type this. Use REA to understand how Obsidian’s quick switcher ranks and matches note titles as you type. [music] Show me the evidence, then build a small web page that reproduces the same matching and

02:10 ranking over these 30 note titles. And it starts strong. Claude loads the REA skill and REA takes the whole bundle apart. 3.9 million syntax nodes, 1,600 findings, 42 message channels between the window and the main process in about a minute. I’m thinking, [music] this is going to be easy. Then it fell flat. Claude asked REA to trace the word switcher through that map, and the trace came back empty. Zero matches. REA’s map doesn’t index every string in a minified bundle. And for a second, I thought, here we go, another

02:45 report that looks great in the readme. But watch what Claude does next. It doesn’t give up, and it doesn’t make something up. It unpacks the bundle and reads the code itself. 3.7 megabytes of minified JavaScript, one byte offset at a time. And it finds the chain. The command is switcher open bound to command O. It runs a fuzzy search. Your query gets lowercased and split into words. Every word has to appear in the title in order. If that fails, it tries letter by letter.

03:16 Then it scores. Every match starts at zero and loses points. One point for every extra chunk it had to highlight. A tenth for every word that started mid-word. A little for how spread out the matches, and a little for how far in it starts. Highest score wins. Claude built the page, and then came the part I didn’t ask for. It ran Obsidian’s original functions and its own copy side by side over all 30 titles and 4,915 queries. Same rank lists, same scores,

03:47 same highlights every single time. Here it is. I type rev weekly review first, then design review, then renovation. I type DN. Reading list beats daily note. That’s a real Obsidian quirk, and the copy has it, too. 4 minutes, 11 seconds. If this is useful, hit subscribe because it’s free. And if you want to back the work, hit join here or the supporter button on Facebook. Now, the question that got Claude stopped.

04:18 Test two is the one that matters if you ship software. Forget copying a feature. I wanted to know what Obsidian’s window is allowed to ask the system to do behind your back. So, I typed this. r e a found all 42 channels in seconds with the exact spot in the code where each one is handled. Claude started reading the handlers and writing the [music] map, and then this happened. Interrupted. Stopped by a safety classifier. Opus 5.5’s safeguards flagged the session. A note explains it can sometimes flag work that isn’t about

04:51 hacking at all, and offers to switch to Opus 4.8. Think about what that means. The model itself decided that mapping what an app can do to your machine looked too much like an attack. So, I switched to Opus 4.8 and ran it again. This time it finished. 42 channels in four groups. 15 touch the file system. Copy a file into the app folder, write an icon, move something to the trash, create or rename a vault folder. One, vault request URL. Lets the window send any web request through the main process. Any address, any headers,

05:24 any body. And one, register CLI. Runs a shell command to put Obsidian on your path. Every line points at the exact spot in the code. To be clear, none of that is a flaw. Obsidian needs those doors to be a note app. The point is that the full map of what an app can do to your machine with the code locations now takes one prompt. And the same map exists for every app you ship. And Obsidian is the easy case cuz JavaScript apps carry their code with them. REA’s own case studies go further.

05:57 These are from area.tools, credit to the REA team. Notion, they followed copy and paste from the page through the preload into the main process out to your clipboard. Four pieces of code, one channel name, DXBall, a Windows game from the ’90s. They followed a sound call into the function that decides which speaker a sound comes from. Rewrote it in C, and the rebuild reproduces all 63 bytes of the original compiled function. And the one that made me sit back, Windows Calculator. Why does 200 + 10%

06:30 give 220? Before REA, you read assembly. With REA, you ask your agent. Every closed app on your computer just became readable. Three things before you run this. One, it does not give you the original source code back. Nobody can. You get pseudo code, maps, call paths, and evidence, and the agent still has to think. You saw that in test one. Two, the evidence files are big. Obsidian’s was over 500 megabytes. REA

07:01 saves it and hands the agent a summary, which is the right call, but plan for the disk space. Three, looking inside software to learn how it works is normal engineering, and in many places it’s allowed for learning and compatibility. Copying someone’s code into your product or breaking a license check is not. Read the license of the app you pointed at. I put every prompt from this video into a free one-page sheet. The setup line, both Obsidian prompts, a website prompt, a native binary prompt, and what to expect back from each. Comment inside,

07:33 and I’ll send it to you. New to Cloud Code? My beginner guides are in the description. So, here’s my question for you. Would you point this at your own app before someone else does? Tell me in the comments. Follow along on Instagram and Facebook for bite-size AI. And the deep dives live here on YouTube.