Android’s sideloading adjustments pushed me to construct an APK installer

I don’t have an issue utilizing ADB through the command line each every so often. However between remembering the place I put the executable, opening a terminal window, connecting a tool, and pasting within the APK file path, it’s a little bit of a chore. Booting up a small app on my PC to deal with a lot of the legwork could be a lot quicker.

Sure, there are already choices available on the market, resembling ADB AppControl, EasyADB, or QuickADB. Nonetheless, I wished to construct one thing myself that particularly focuses on the APK set up workflow — wi-fi debugging pairing, file/listing looking, and bundle unpacking — somewhat than a full ADB management suite.

Whereas I’m a reasonably succesful programmer, between work and two children, I don’t have the time to jot down an app completely from scratch. So I assumed I’d give AI-assisted “vibe coding” a attempt. This had the additional advantage of letting me check out a brand new programming language I wasn’t accustomed to: Go.

Why Go? It’s a easy language that’s an ideal match for a compact undertaking like this (no want for Rust), but it simply compiles right into a small, cross-platform executable (sorry, Python). Plus, I wished to construct a cool-looking terminal app utilizing the gorgeous Bubble Tea TUI framework that I’d not too long ago come throughout. With the design in thoughts, I set off to search out an AI agent to assist me.

I’ve dabbled in AI instruments sufficient to know that merely asking one to construct an software for you is a raffle. Costly fashions from Anthropic and OpenAI can one-shot easy duties, however you usually find yourself in an exhausting back-and-forth to resolve bugs once you let an agent free on a fancy undertaking. To work together with ADB and construct a terminal-based UI, I made a decision I wished a extra hands-on strategy.

I wasn’t about to spend a whole lot of {dollars} on Opus 4.6 for this little undertaking. Even with cautious planning, there’s at all times a threat that fashions get caught in an costly construct loop or hit an inescapable drawback. As an alternative of asking Claude or Codex to deal with the duty themselves, I paired my standard IDE with OpenCode and its very inexpensive $10 Go subscription. That gave me entry to MiniMax M2.7 and the very spectacular GLM 5 to make use of alongside my slower human mind and keyboard-worn fingers.

I laid out the naked bones of the construction myself—folders, a couple of placeholder features, a .gitignore, and so forth—and wrote a fast AGENTS.md to stipulate the undertaking’s necessities, objectives, and coding specifics. Then, somewhat than asking AI to complete the entire thing, I moved slowly by the phases I knew I’d want: planning, constructing, reviewing, and tweaking every step alongside the best way. I extremely advocate utilizing model management, like Git, so you’ll be able to undo main AI errors and climb again out of inevitable rabbit holes.

First, I bought a fundamental ADB wrapper working in Go, permitting me to run instructions like adb gadgets to search out related telephones after which utilizing adb set up to push APKs to them. Subsequent, I expanded the working instructions right into a fundamental terminal interface to step by the method. That was fairly easy, however not probably the most user-friendly expertise.

Lastly — and positively probably the most arduous a part of the back-and-forth — was establishing the UI move. It didn’t assist that I used to be studying a brand new language and framework on the identical time, however by instructing the AI agent to reveal theming by a configuration file, I used to be capable of make guide tweaks myself. That saved time (and credit) by avoiding the necessity to burn assets on each small element.

Ultimately, the mixture of human planning, AI constructing, and tag-team tweaking produced a strong little app in simply three days of off-duty work. I definitely wouldn’t have been capable of write the ~3,000 traces of code on my own in that point. Not dangerous in any respect for $10, and I’ve loads of credit left for an additional thought.

It’s a compact 7MB executable — no set up required. It may possibly mechanically obtain ADB from Google in the event you don’t have already got it, or use an current PATH set up. The app launches ADB when wanted and shuts it down on exit. You possibly can pair with debuggable Android gadgets over USB or Wi-Fi, browse recordsdata utilizing a built-in file explorer, and set up APKs instantly. It additionally helps unpacking and putting in app bundles, together with APKM, XAPK, and APKS recordsdata.

The excellent news is that I bought a working app, and it’s genuinely helpful. If there’s one downside, it’s that I didn’t be taught as a lot Go as I’d initially hoped. By asking AI to unravel issues for me, I missed out on deeply studying the language’s syntax and patterns. I’ve picked up some understanding from reviewing the code and following the logic, however with out doing a lot of the problem-solving myself, I nonetheless wouldn’t name myself competent in Go.

And that’s the true threat with vibe coding: not totally understanding how the code works. I noticed a few innocent bugs and even a zipper slip vulnerability whereas reviewing AI-generated adjustments, however I can solely think about the dangers when constructing one thing involving encryption, API keys, or delicate credentials. My recommendation to anybody attempting vibe coding is to place within the effort to plan and perceive not simply how your app ought to work, but in addition which instruments you want, the way to handle the undertaking, and what generally goes improper. That method, you’re higher positioned to confirm AI output when it inevitably misses the mark.

If you wish to give my app a attempt, you’ll find releases and the supply code on the Tiny APK Installer GitHub repo. I haven’t had an opportunity to check the Mac or Linux builds but, so I’m trusting that Go compiled every little thing appropriately. In case you spot any bugs, be at liberty to open a difficulty — and perhaps curse AI for making it this simple to construct (and break) issues.