How to build your own version of an app you pay for,with AI

If you've ever paid every month for an app on your computer that you barely open, you can now build your own version with AI. On Monday I open-sourced mine, a Mac screen recorder called Rushes, and these are the three decisions that made it work.

Atlassian's help page says Loom won't let you download your own videos on its free plan. Downloads start with the Business plan. On Saturday evening I needed to record myself talking over my screen and keep the file, so I could edit it.

So I had my AI agents build me a recorder that saves my screen, camera and audio as separate files for my editor, all starting at the same instant. Claude Code and ZCode did the research and the plan that night. When I hit my ZCode usage limit on Sunday afternoon, Codex picked it up and kept building. The agents named it Rushes, a film term for the day's raw takes.

My agents wrote the code. I made three decisions you can copy.

Start from the free versions that already exist

Search GitHub for "open source" and the name of the app you pay for. My agents found nine free screen recorders and studied them before writing any of the app's code.

Ask your agent to check each one's license before it copies anything. Three of the nine, Cap, Recordly and Screenity, use copyleft licenses, which can require anything built from their code and shared with others to use the same license. Rushes would have had to take their license if we'd copied their code, so my agents only took notes on those three and built from the notes. The other six, including Kap and OpenScreen, use permissive licenses like MIT, which let you reuse code if you credit it. Rushes reuses two small files from that group, each with its credit and license attached.

That split, where one side describes what the software does and the other builds from the description, is called a clean room. I wrote about where its legal lines are in my last post.

If nobody has built a free version yet, there's a tool for that now

REA, short for Reverse Engineer Anything, lets an AI agent look inside an app on your computer, and even run it, to explain how a feature works. It's free under the MIT license and one of the 60 most-starred projects on GitHub. I haven't tried it yet.

If you do, three cautions:

• Install it only from the original. It lives on its creator's own GitHub page, morluto/rea. Copies have popped up in the past week, and one offers a download you double-click. The original has no files to download; it installs with a terminal command.
• Check the app's terms first. Adobe's, for one, ban reverse engineering. REA's own README says it's for lawful uses, such as analyzing software you own or are authorized to examine.
• Build from notes, not from the app's code. If the terms allow it, have your agent write down what the feature does, then build your version from those notes without copying any of the app's code.

Build the version you want, not a copy

I wanted the best features from all nine, including separate files for my editor. My agents wrote that into a plan, and before any code existed, a second AI model, Kimi K3, graded it 8 out of 10.

It caught a promise the plan couldn't keep: a share link 2 seconds after you hit Stop, when saving the recording could take up to 5. The fixed plan stopped making the link wait for the save. It scored 9, and the build started after that.

Use it for real before you trust it

The agents told me Rushes was installed on Sunday night. At 9:48 my first recording failed the moment it started. So did the second.

Within the hour the agents had a test recording working. When I exported it, I got one combined video, and I couldn't find the separate files I'd built it for. They were behind a menu option.

None of that showed up until I used it for the job I built it for.

Try Rushes

There are good free recorders already. Cap keeps separate tracks too, but its free download is for personal use; that limit doesn't apply if you build Cap yourself from its source code. I wanted one I could use freely, so Rushes is free for anyone, even for work, under the Apache 2.0 license.

After you hit Stop, you can trim or split clips, edit by deleting words from the transcript, and export one finished video or the original tracks for your editor. Editing never touches the original files, and transcription runs on your Mac.

It's early. There's no installer yet, so for now you'd have your AI agent build it on a Mac running macOS 26.4 or later. If you try it, tell me what breaks.

Next time a subscription renews, check whether someone has built a free version. If nobody has, you might be the one.

Which app you pay for would you build your own version of first?