Reverse Engineer Anything (REA): Can AI Rebuild an App Feature?

Reverse Engineer Anything, or REA, helps an AI coding tool study how a piece of software behaves and rebuild a feature. It is a way to investigate one function you can see, not a shortcut to a finished copy of an entire product. I had not tested REA myself when I published this article on 18 October 2026, so the examples below are credited to the people who ran them.

What is Reverse Engineer Anything?

REA is a free project on GitHub that connects AI coding agents to software analysis tools. Reverse engineering means starting with a finished app and working backwards to understand how it was built. For a non-developer, the useful idea is that the app can provide some of the detail you might struggle to describe in a written brief.

The project describes a sequence: recover readable material, investigate how the feature works, then recreate it in your project. It reports evidence, confidence and limits for its findings. I use “seen, inferred and unknown” as a plain-English way to sort that information; those are my labels, not official REA labels.

Seen means there is direct evidence in the app. Inferred means the AI is making a reasoned guess that needs testing. Unknown means the necessary information is not present in what it examined. Asking for those three piles makes it easier to catch a confident guess before you build on it.

What have other people built with REA?

Nick Saraev compared two AI-built versions of Apple's Preview app, one made with REA and one without. In his video, the version without REA finished first. The REA version took a few minutes longer, looked closer to Preview and drew more smoothly. That is one creator's test, not a benchmark.

Nate Herk studied a Loom-style recorder. His rebuilt version used a movable camera bubble and a countdown, while saving recordings locally instead of on Loom's servers. It copied a feature's behaviour without using Loom's service or branding.

These examples suggest a useful starting point: choose one feature, then compare the result with the original. They do not prove that REA can clone any app or that every platform is supported equally well.

Where might a non-technical person use it?

Start with the one action you miss in a tool you pay for. A small local utility, such as a recorder or PDF mark-up feature, is easier to inspect and maintain than an entire app. A cloud service can hide its most valuable behaviour on its servers, where a local analysis tool cannot see it.

My own idea is a meeting note-taker that I control, but I had not tested that use. A local interface could be studied; the transcription and summarising service would still need a separate solution. That distinction matters when a demonstration looks more complete than the underlying build.

Is REA safe to install and legal to use?

The official repo says it is intended for lawful research and software you own or are authorised to examine. Some analysis tools it connects to require separate licences. Software terms and intellectual property rules also matter before you take apart an app or sell a rebuilt feature. Check the licence and get legal advice for a commercial use.

Use the official project source rather than a lookalike download. Remember that the AI tool you connect may receive information REA finds in an app. Ask it to show a setup plan and the access it needs before installation.

Should you try REA now?

If you already use Claude Code or Codex, pick one narrow feature and ask the agent to show evidence, guesses and unknowns before it writes code. Test its behaviour against the original. For a visual website feature, a plain AI agent may already get close from screenshots; bring in REA when details of the behaviour are missing.

I have not run REA myself. My verdict comes from the project documentation and published demonstrations, which have their own limits. A rebuilt feature also becomes your responsibility to maintain.

The full Substack article includes the wider story, screenshots, the demonstrations and a starter prompt. Begin by writing down the single feature you would miss most if you cancelled a software subscription.