Nick Saraev tried the viral Reverse Engineer Anything (REA) repo on a familiar target: Apple's Mac Preview photo viewer. In his video comparison, he asked one Claude agent to rebuild Preview with REA and another to do it without the library. The plain agent finished first. The REA-assisted result looked closer to Preview to Nick and seemed quicker when he drew an annotation. That is an interesting outcome, but it is one informal demonstration, not evidence that REA can reproduce arbitrary apps.
REA makes established analysis tools available to coding agents through MCP and CLI. In Nick's test, its deeper inspection appeared to help the agent match some Preview interactions, while adding setup and waiting time. The video does not provide a repeatable speed, fidelity, or cost benchmark. Use the comparison to design better tests, not to conclude that closed-source software is now fully open.
Watch Nick Saraev's REA Test
Credit: The setup, side-by-side Preview experiment, and conclusions are Nick Saraev's. The REA maintainers publish the underlying repo and documentation. The title says “24 Hrs,” but the published video does not provide 24 hours of timed benchmark data.
What REA Adds to an Agent
Nick's explanation is directionally right: the agent asks REA to inspect a target, REA passes work to the appropriate analysis tool, and the agent uses the returned evidence to plan a new implementation. The project's own description is more precise. REA is an interface, not a newly invented decompiler. Native analysis can use Ghidra, Hopper, or IDA; other workflows inspect JavaScript and Electron apps, websites, managed code, and additional supported artifacts. The requirements vary by target.
The useful output is not “the original source code.” It is evidence such as pseudocode, strings, call relationships, modules, screenshots, or observed behavior, plus unresolved questions. The agent may then write a new implementation and test it. Our earlier REA guide goes deeper on a stripped-binary test and the project's verified DX-Ball case study.
The Mac Preview Comparison
| Stage | What Nick shows | What we can conclude |
|---|---|---|
| Target | Mac Preview displaying an image, with zoom, info, sharing, markup, rotation, and save controls. | The experiment tests a visible slice of one desktop app, not all Preview functionality. |
| Two attempts | One Claude session uses REA; another is asked to recreate Preview without REA. | A useful side-by-side, but the prompts, inspection inputs, and resources are not shown as controlled equivalents. |
| First finish | The plain-agent version finishes while the REA-assisted version is still running. | In this run, the plain route reaches a previewable result sooner. |
| Manual check | Nick compares the windows and tries zoom, sharing, drawing, rotation, annotation, and save. He prefers the REA-assisted match and calls drawing more responsive. | There is a creator-observed quality difference on those interactions, but no measured fidelity or latency score. |
The trade-off is the story: deeper evidence may improve a replica of selected behavior, yet the first usable result may arrive faster without the extra analysis. The video does not show repeated runs, a full feature checklist, identical prompts, cost logs, or independent reviewers. It also does not establish whether the new app handles large files, color management, accessibility, edge cases, or future macOS versions.
What the Hype Gets Right and Wrong
At 05:26, Nick makes an important distinction: Ghidra, Hopper, and IDA predate this social-media moment. REA's contribution is the agent-facing workflow over such tools, not the invention of reverse engineering. The project's fast-growing star count measures attention, not reliable reconstruction accuracy.
Nick then argues that agents will make many software surfaces easier to reproduce and some narrow app ideas harder to defend. That is a strategic prediction, not a measured finding from the Preview test. The same video shows why the distinction matters: a plain agent already made a plausible Preview-like interface without REA, while the REA-assisted version was not proven equivalent to Apple's product. Neither path grants rights to reuse Apple's code, art, name, or protected behavior.
There is also a practical security boundary. REA's runtime-capture documentation says a target can run with the user's own permissions. Review setup changes, isolate unfamiliar binaries, and decide what findings may be sent to an external model provider. For any commercial target, obtain permission and a legal review appropriate to the jurisdiction and purpose.
A Better First Experiment
- Use a program you own. Pick a small internal utility with source and tests you can keep hidden as the answer key.
- Freeze the target. Give both agents the same written feature list, sample inputs, time budget, and acceptance tests. Record whether the REA agent also gets a binary or source map.
- Measure both routes. Log setup time, time to first usable build, working features, crashes, output quality, and human repair time. Avoid choosing a winner on appearance alone.
- Check the unknowns. Test invalid files, permissions, save behavior, accessibility, and performance on a larger input. Mark inferred behavior separately from confirmed behavior.
- Stop before production. Have the owner review code, security, licensing, and real-user needs before replacing any existing workflow.
Jump to a Segment
| Time | Topic | Time | Topic |
|---|---|---|---|
| 00:00 | REA goes viral | 01:08 | How REA works |
| 02:44 | Preview rebuild | 04:46 | Matching the app |
| 05:26 | What is not new | 07:27 | What comes next |
Turn an Authorized Comparison Into a Service
A useful offer is a fixed-scope migration feasibility audit for an owner whose old utility is hard to maintain. Sell the tested behavior map and decision, not a promise to clone a competitor. This prompt works in a general AI assistant and does not require REA to start.
Audit one owned workflow
Compare an evidence-led rebuild with a plain prototype.
Sources and Limits
Creator: Nick Saraev's video and channel. Project: REA repository and README, official guides, and Ghidra source. Further analysis: our earlier REA article covers a controlled stripped-binary test, evidence standards, and legal/security cautions. Nick's impressions of visual similarity and speed are attributed to him; no independent benchmark files accompany the supplied transcript.