Google AI Studio Now Puts Prompt-Built Android Apps on Phones

From a typed request to an app on your phone

Google has closed the last big gap in its plan for prompt-driven app creation. As of October 13, anyone with an ordinary Google account can describe an app in Google AI Studio, have it built, and install it on their own Android phone through a private section of Google Play. The apps don't appear in the public Play Store catalog.1113 Users don't need to pay for a Play developer account, handle APK files, or open a developer console.15 Building is free.11

Coverage this week presented it as a consumer launch, but the October update finishes work Google began in the spring. At Google I/O on May 19, 2026, AI Studio's Build mode gained an Android option. It turns a prompt into a Kotlin project, runs the result in an emulator inside the browser, and supports installation over USB.132 What remained was the step from the browser to the phone. Before October, getting an app onto a handset meant either turning on developer options and connecting a cable, or paying for a Play developer account.13 Sam Bright, Google Play's VP and general manager, summed up the change as moving from "there's an app for that" to "there's an app for me."13

What the machine actually writes

The output is real Android code, and that sets this apart from most vibe-coding tools. Many of them produce web apps, and as The New Stack noted at the I/O launch, mobile development had been left behind while prototypes ran on localhost in a browser.2 AI Studio's Antigravity Agent writes a full Gradle-based project instead of a web mockup.14 It uses Kotlin, Jetpack Compose and Material 3, with a single-activity structure, ViewModels and the usual manifest and resource files.1413 Google says the technology is the same one behind Gemini in Android Studio.2

The code is visible and editable inside AI Studio, and users can download the project as a ZIP file to keep working in Android Studio.144 One analysis compared the result to what Android Studio's own new-project wizard produces. It also pointed out that Google's SummitPack demo used Room for local storage, Retrofit for Gemini calls and an MVVM structure, all libraries a human Android developer would likely choose.13 Users can keep refining the app with chat prompts or edit the code directly.1

People started using it quickly. Google's Logan Kilpatrick said more than 250,000 Android apps were built in the first week after the May launch. He estimated that more than 99% of those builders had never made an Android app before.13

Where the reporting diverges

Outlets agree on the basic mechanics but differ on several details.

Availability. Techweez and a DEV Community write-up say personal apps launch for adults aged 18 and over in the U.S. and a few other markets.1517 Yahoo Tech says the feature is now public after months of pre-registration and beta testing.4 The more cautious account is likely closer to how the rollout works in practice: generally released, but limited by region and age.

Hardware access. At I/O, The New Stack described native access to sensors, GPS, Bluetooth and NFC as the main advantage over web apps.2 Later coverage makes a distinction. Generated apps can use the camera, GPS, accelerometer and Bluetooth on a real phone, but the browser emulator can't simulate camera capture, NFC, Bluetooth or Google Play services.413 So developers can only test some features once the app is on an actual device.

Safety review. This is the largest disagreement. Techweez reports that private apps still have to follow Play policy but are not reviewed by Google while they stay private. Safety checks apply only if a builder makes the app public.15 A hands-on guide on another site claims every bundle gets automated malware scanning and that personal builds are restricted to basic device APIs.6 The more detailed claims in that guide don't appear in other coverage, so readers should treat them cautiously. For now, the best-supported reading is that a private app gets little review before it runs on its owner's phone.

The limits that define the product

Read closely, Google's documentation describes a narrow tool. Native Android projects in AI Studio run on the client only. Firebase integration, Workspace APIs, secrets management and multiplayer features are available for web apps but not for these native builds.14 The builder supports only single-module, single-activity Compose projects. Java, XML layouts, NDK code, Wear OS and Android TV are not supported.144 Google's documentation updated in August still listed Firebase, GitHub export, Wear OS and TV as unsupported.13 That was notable because Firebase had been on the roadmap since I/O.2

The Play integration for developers has a clear limit too. Users with a Play developer account, which costs a one-time $25 fee, can have AI Studio create the app record, package the bundle and upload it to an internal testing track.2 Google's documentation says AI Studio publishing stops at internal testing, and production releases still go through Play Console.14 Google has not announced a badge or storefront category for AI-generated apps. Anything that reaches the public catalog goes through the normal publishing process.14 Sharing a personal app beyond your own devices still requires the standard developer process for now.4

Costs are low at the start. The free tier has a modest daily quota, and content created on that tier can be used to improve Google's products.13 Paid AI Pro and Ultra plans raise the quota, and apps that call Gemini are billed at Gemini API rates once they move beyond the free tier.13

Why this matters for AI-first software development

The bigger significance is what Google chose to make easy. Generating code is now cheap across the industry. The hard parts of mobile development were the toolchain, signing and distribution. Notebookcheck pointed out that Android Studio's installer alone is about 1.4 GB, while AI Studio needs only a browser.7 With the private Play library, Google has also removed the last step of getting the app onto a phone. In practice, the company is turning its app store into a distribution channel for software made by one person for their own use.

ZDNET argued that the appeal lies in very small, personal tools: a habit tracker, a unit converter, a timer, or replacements for single-purpose apps full of ads.11 Techweez added another reason. Google has made sideloading harder over the past year, which pushed many hobbyists toward web apps even when a native app would have been better.15 A Google-approved route for personal native apps solves that problem and keeps users inside Google's ecosystem. One commentator described it as a private sandbox built into the Play client.6

The workflow that's emerging has stages. Builders start in AI Studio, test the idea on their own phone, and export the ZIP once they run into the single-activity or no-Firebase limits. After that, they can use whichever local coding agent they prefer.13 Paul Thurrott's account of a weekend spent vibe coding shows both sides. He reported good results in AI Studio, then hit garbled output and an apparent quota limit with Gemini inside Android Studio, and finished the project by adding Anthropic's Claude Code.5 The tools work well for a first draft, but finishing a project still takes switching between tools and some trial and error.

The quality question doesn't go away

Most of the more careful coverage agrees that a prompt produces an app project, not a finished product. Before wider release, generated code still has to be tested for crashes, accessibility, permissions, privacy, security and compatibility.14 An Android-focused critique on DEV Community explains why mobile is especially hard for AI. Models trained on older data may use deprecated APIs such as AsyncTask, misuse coroutine scopes so work leaks in the background, overuse Kotlin's !! operator, and write Compose code that triggers unnecessary recompositions.10 A hands-on tester reported one of these bugs firsthand: a counter reset on every rotation until a follow-up prompt fixed how the state was stored.6

This is why Google's restrictions make sense, even though they get less attention than the launch. Single-activity, client-only apps running on the owner's phone carry the lowest risk. If a bug only affects the person who asked for the app, the cost of AI-generated mistakes stays small. The harder questions will come with the sharing features Google has promised. The DEV Community write-up describes plans for secure sharing and, later, standard store review.17 Once vibe-coded apps start spreading between friends and families, the lack of review for private apps will draw more scrutiny.

For now, the takeaway is that Google has made building an app for personal use about as easy as writing a request. Building software other people can rely on is still the same engineering work it was before.