REA: The GitHub Repo That Lets AI Agents Take Apart Any App
REA, short for “Reverse Engineer Anything,” has been sitting at the top of GitHub Trending for days. Miles Deutscher of AI Edge calls it the most insane repo he has seen. The idea is that an AI agent can open up an app, website, or game, work out how a feature actually functions, and use that to build your own version. His video walks through setup, a five-step workflow, and an agent swarm that runs the whole thing automatically.
Why an AI agent needs REA
You might wonder why you can’t just send Claude a screenshot or a URL. Deutscher’s answer is that Claude reads, reasons, and writes code well, but on its own it can’t open an app on your machine and digest what is happening underneath: the code, the features, the behavior. REA is the bridge. It connects whichever agent you use (Claude, ChatGPT, or another model) to tools that inspect native binaries, JavaScript and Electron apps, .NET assemblies, and websites.
Two details stand out. The analysis runs locally on your computer. And REA reports its limitations, the evidence behind each finding, and a conclusion, so the agent isn’t guessing.
Who it’s for
Deutscher says to think beyond developers:
- Designers can study how a site like Apple’s achieves a design effect.
- Game builders can learn a mechanic they like.
- Product builders can point the agent at something like Notion’s drag-and-drop CRM board instead of explaining it with screenshots.
- Entrepreneurs and agencies can build custom dashboards, finance trackers, or client interfaces.
He builds trading and agentic finance products and runs a media company that makes funnels and websites, and he says REA has helped on both sides.
The legal warning
Before going further, he says not to rip off someone’s app and sell it. There are IP and legal limits, and he stresses doing everything legally. He repeats near the end that you should check the terms of service and the laws in your jurisdiction.
What people have already built
REA’s own examples include a recreation of Notion’s clipboard and the old game DX Ball. Online, people have posted a Roblox recreation and even a GTA-style game with Sonic in it. Deutscher finds the most practical use is picking individual features and workflows you like and rebuilding them inside your own stack. He also sees a sovereignty angle: you can run the whole investigation flow locally, so you aren’t dependent on a service provider.
The five-step framework
His framework gives five ways to use REA:
- Look. Map the general architecture of a feature.
- Understand. Break down exactly how it works.
- Plan. Turn that understanding into a usable spec.
- Build. Use the spec to create your own version of the app, tool, game, website, or workflow.
- Iterate. Improve it over time.
He demos the sequence with Obsidian, the markdown note-taking app known for its graph view.
Step 1 identifies the target. The agent works out that Obsidian is an Electron app, picks the right REA tool (the JavaScript analyzer, so no external tool like Hopper is needed), lists what each component is built with, and notes where to dig next. He suggests similar prompts for other targets, such as asking how the animated gradient on Stripe’s homepage works.
Step 2 is an investigate, or X-ray, prompt. He asks for a plain-English, step-by-step account of what happens when a user opens the graph, what data it uses, and how the pieces connect. The agent turns that into a graph of its own. His point is that without REA an agent guesses how something works, while with it you get precision. That precision lets you one-shot the build and saves tokens, especially on complex builds with 3D graphics or animation.
Step 3 produces a full spec: the data needed, requirements, look and feel, edge cases, errors, and test cases.
Step 4 reads that spec and builds. In the demo, the agent recreates Obsidian, including a working graph and per-file breakdowns.
Step 5 is an upgrade prompt. It names three things the new version does well and three weaknesses or missing pieces, then fixes them.
The agent swarm
The approach he likes most is a swarm. You give an input to a lead agent, which sends out multiple scouts to investigate. A spec writer, a builder, and a tester follow, with the tester feeding results back to the builder. The swarm produces a final report and can loop and self-improve. He expects people to automate this so agents test products and features overnight and bring back improvements. As he puts it, it has never been a better time to be a vibe coder.
Practical uses
- Recreate features from apps you like, or build a custom operating system from them.
- Build custom UIs for clients if you run an automation agency.
- Own your software and stop paying around $30 a month for a tool when you only need one or two of its features.
- Build digital products and research and iterate on a market faster.
- Verify claims made by lesser-known vendors.
- Build tooling, such as custom connectors, automations, and MCP bridges.
He also points to others doing this. Eric Tech cloned a Wispr Flow-style voice transcription app so he owns his transcripts, and another creator, Nate, reverse engineered Loom to build screen-recording software.
A note on models
Some prompts, he says, may hit guardrails on models like Anthropic’s Opus, even for legitimate work, and browser use can be blocked or downgraded. His advice is to test on Claude and GPT first and see whether you hit limits. If you do, he mentions Grok, which tends to block less, or local Qwen models, depending on how much RAM you have. Whichever model you use, the legal caveats above still apply.
Getting started
Deutscher keeps the full setup prompt and all the prompts from the video in his free Skool community, so you can paste them into a terminal, Claude Code, or Codex. His advice is to try it yourself, starting with something you’re allowed to study, and he offers a follow-up video that reverse engineers one app from start to finish.
