Google AI Studio now generates complete Android apps in Kotlin with Gemini 4 Hydrogen

- Reading
- 5 min
- Code
- 1
Since November 17, 2026, the Google AI Studio Build agent provisions Firestore, Google login, and secure Gemini calls directly in Kotlin, and publishes the app to the Play Store without a manual keystore.
What's changed for Android developers
Until recently, generating an Android app in Google AI Studio meant being stuck with local session storage. No persistence across devices, no real login, no secure way to call the Gemini API without exposing the key inside the APK. It was prototyping, not a product.
On November 17, 2026, Google announced that the AI Studio Build agent, running on the Gemini 4 Carbon and Hydrogen models, now assembles the entire cycle: database, authentication, protected Gemini calls, and publishing to Google Play, all from a natural language prompt. The final code comes out in idiomatic Kotlin, ready to open in Android Studio if you want to keep working on it yourself.
From prompt to Kotlin: what the agent builds on its own
When the prompt describes an app that needs persistent storage, login, or generative AI features, the Build agent detects this on its own and asks whether it can provision the backend. Once confirmed, it:
- Creates a Cloud Firestore database and generates Firebase Security Rules tailored to the app's data model, including offline cache that syncs when the connection comes back;
- Configures Firebase Authentication with Google Sign-In and per-user authorization rules, so no one can read or write someone else's document;
- Wires the Firebase Bill of Materials (BoM) into the project and writes the Kotlin code that talks to Gemini and to Firestore.
The part that matters most to anyone who's already suffered through this in production is the call to the model. Hardcoding an API key inside an APK is practically an invitation to reverse engineering. The agent solves this by configuring Firebase AI Logic, which keeps the credential on the server, and Firebase App Check reinforced by Google Fraud Defense (formerly reCAPTCHA Enterprise), which validates whether the request really comes from your app, both in the emulator and once published on the Play Store.
The code that comes out the other side
Google's announcement includes a sample snippet of what the Build agent generates for a fictional workout app that creates routines with Gemini and saves them to Firestore:
// Generated by the Google AI Studio Build agent using Firebase AI Logic and Cloud Firestore
class WorkoutRepository(context: Context) {
private val generativeModel by lazy {
Firebase.ai(backend = GenerativeBackend.googleAI())
.generativeModel("gemini-4-hydrogen")
}
private val db = FirebaseFirestore.getInstance(
context.applicationContext.getString(R.string.firestore_database_id)
)
suspend fun generateAndSaveWorkout(goal: String, uid: String) {
val response = generativeModel.generateContent("Create a 20-minute workout for: $goal")
val workout = hashMapOf(
"ownerId" to uid,
"goal" to goal,
"plan" to (response.text ?: ""),
"createdAt" to FieldValue.serverTimestamp()
)
db.collection("workouts").add(workout).await()
}
}This isn't toy code: it uses suspend fun, handles the nullability of response.text, and separates generativeModel as a lazy property, exactly the pattern an experienced Android developer would write by hand. This matters more than it seems: if the generated code were disposable, the tool would just be a prototype generator. Here, the idea is for it to become the project's real starting point.
HD emulator, audio, and screen capture for the agent
The second front of the announcement touches the visual iteration loop, historically the bottleneck of mobile development in any browser-based tool. The changes:
- Faster emulator boot and streaming at up to 1200p, with an HD toggle to inspect fine spacing in Jetpack Compose;
- Real-time audio directly in the browser, which unlocks testing voice companions with the Gemini Live API, TTS, and media apps without leaving AI Studio;
- One-click screen capture: you take a screenshot of the running emulator, attach it to the prompt, and ask for the visual adjustment ("tighten the card's padding," "fix the contrast") instead of describing it in text;
- Custom fonts applied to the Compose theme, with the official Android CLI behind the scaffolding, which reduces the risk of the generated project coming out in a way that doesn't compile outside of AI Studio.
Publishing to the Play Store without a manual keystore
The part that historically trips up anyone who just wants to test on their own phone is installing the APK outside Android Studio: enabling USB debugging, managing a signing keystore, opening a Play Console account. With Google Play's direct integration into AI Studio, you can set up a Play account in a few clicks in the publishing panel, and AI Studio itself compiles and signs the Android App Bundle (.aab) behind the scenes. The app shows up in your personal Google Play library, installable on your phone without a USB cable.
What's still left open
It's worth separating what the announcement solves from what it doesn't. It's good to be skeptical here, even though the advance is real:
- Firebase lock-in: the generated backend is Cloud Firestore and Firebase Auth by default; migrating to Postgres, Supabase, or a custom backend later means rewriting the data layer, not just swapping configuration;
- Nothing replaces reviewing Security Rules: the agent generates rules "tailored to your data model," but a poorly written Firestore rule is one of the most common causes of data leaks in real apps, and that remains the responsibility of whoever reviews the code before publishing;
- The announcement doesn't mention automated testing, CI/CD, or UX quality review beyond the visual, which remain manual work after the agent delivers the first version;
- It's Android first: there's no mention of iOS or cross-platform support in this flow, which keeps AI Studio as a prototyping and mobile MVP tool, not a replacement for your cross-platform stack.
How to try it
Google suggests starting by opening Google AI Studio Build, selecting the Android chip at the top of the workspace, and trying a prompt like "Build an Android workout coach app that uses the Gemini API to generate personalized daily routines, reads cues aloud with audio, and saves workout history across devices using Cloud Firestore and Google Sign-In". Another example from the announcement itself asks for a family recipe app with serif typography, Google login, and a shared shopping list in Firestore.
The sensible path for anyone who wants to evaluate this carefully is to first ask for something small, with just one data model and one screen, inspect the generated Kotlin and the Security Rules before publishing, and only then scale the prompt up to an app with multiple Firestore collections. Since the feature just launched yesterday, there's no independent report yet of limits or production bugs, which is the first thing to watch over the coming weeks.
Translated from the Brazilian Portuguese original · Read the original
OpenAI launches GPT-6.2 Astra with price cut and GPT-6.2 Sol for everyday use
GPT-6.2 Astra arrives 20% cheaper than its predecessor and matches Claude Fable 5.2 on price; GPT-6.2 Sol keeps the same price as the two previous generations, half the cost of three generations ago. For those running agents in production, the math changes more on cache than on raw tokens.





