Vibe coding in Google AI Studio: build, secure and deploy an app with Build mode

What is Build mode in Google AI Studio?

Build mode is the part of Google AI Studio where you describe an app and Gemini writes it. Google's documentation calls this vibe coding, and says it is meant for quickly building and deploying apps that show off what Gemini can do, such as image generation and the Live API. You see the generated files and a live preview side by side.

The engine is the Antigravity Agent. The docs say the Antigravity Agent, the main AI feature of Google Antigravity, now provides the core of the Build mode experience. In plain terms, it remembers earlier prompts and file states, edits several files at once, and checks its own changes by running them.

Two things changed recently that matter if you read older tutorials. First, Build mode is now full-stack, with a server-side runtime, not just a browser prototype. Second, Google is closing Firebase Studio on March 22, 2027 and points its users here; see the Firebase Studio guide.

How do you build an app in Google AI Studio?

You need a Google account. Go to Google AI Studio and open Build. The docs list four ways to start:

• A prompt. Describe the app in the input box. Add AI Chips to include a capability such as image generation or Google Maps data, or dictate with the speech-to-text button.
• I'm Feeling Lucky. Gemini writes a project idea and a prompt for you.
• Remix from the gallery. Open an app in the App Gallery and choose Copy App.
• Import from GitHub. Use the plus icon in the prompt box, then Import from GitHub.

1. Pick the platform. The platform picker offers web app (the default) or native Android app.
2. Write the first prompt using the template below. Be specific about the one thing the app does.
3. Run it. The files generate, and the preview opens on the right. The Code tab shows what was written.
4. Try the app like a user. Click every button. Note what is wrong in one list.
5. Fix in chat, or point at the problem. Annotation mode lets you highlight part of the UI and describe the change you want.
6. Save a checkpoint to GitHub before you ask for anything big.
7. Publish when the core flow works.

Build a web app called Receipt Splitter.

User: two friends who shared a restaurant bill.
Core flow: upload a photo of a receipt, use Gemini to read the line items, let the user tap which items each person had, show each person's total including tax and tip.

Keep it to one screen at a time: Upload, Assign, Totals. No login. No saved data yet.
Do not add features I did not ask for.
Done when: I can upload a receipt photo and get totals that add up to the receipt total. Show an error message if the photo cannot be read.

That template has the same shape as the build prompts on the board: who it is for, the core flow, what to leave out, and how you will know it works. More on that in how to write an app build prompt.

What does Build mode actually create?

Because the web app has a server, the agent can install npm packages on its own when it needs them. You can also ask for one by name, for example “use axios to fetch data from the external API”. Google's page on full-stack apps in AI Studio describes the same runtime.

How do you add a database, sign-in and API keys?

You ask in chat. The agent writes the code and, for Google services, handles the setup.

Camera, microphone and location access need an extra step. The docs say to list the permissions in the app's metadata.json using requestFramePermissions, for example microphone, camera and geolocation, so that viewers are asked to acknowledge them.

How do you deploy an AI Studio app?

Click Publish in the top right. Every deployment becomes a service in Cloud Run, and Google describes the deploy options in two tiers.

• Who is not eligible for Starter Tier. Google Workspace accounts (paid or free, including Education and Nonprofits), accounts with an active or prior paid Google Cloud billing account, and accounts without enough activity history. Support cannot override this.
• Custom URL. You can pick a subdomain such as your-app-name.ai.studio on first-come, first-served basis. Unpublish or delete the app and the name is released.
• Your key is shared by all visitors. On Cloud Run the app uses your API key for every user's Gemini calls, so your usage and any costs add up as people use it.
• Sharing is not private. Users you share with can see the code and fork it. Share with edit permission and they can change it.

Can you take the code out of Google AI Studio?

Yes, in two ways. You can download the code as a ZIP and open it in your editor. Or you can use the GitHub tab in Settings to create or link a repository and sync both ways. AI Studio pushes your prompted changes with AI-written commit messages, and pulls changes you made locally or that teammates merged. If both sides changed the same file, a Resolve conflicts dialog shows a side-by-side diff.

Two-way sync is the safest way to graduate from prompting to normal development. Keep prompting in AI Studio for the first working version. Then clone the repo and carry on with a coding agent such as Claude Code or Codex CLI when you want tests, refactors or CI. The Code tab also lets you export to Antigravity.

Is Google AI Studio free, and what are the limits?

The Gemini API pricing page states that Google AI Studio usage is free of charge in all available regions. The cost sits in what your app calls and where you host it:

Check the Gemini API pricing page for current free-tier limits before you share an app widely. They change by model.

When should you use something other than AI Studio?

AI Studio is strongest when the app uses Gemini, Google sign-in, Firestore or Workspace data, because the setup is automatic. It is weaker if you need a non-Google stack, a team editing live, or heavy refactoring over months.

• You want a broader set of prompt-to-app tools. See the best AI app builders.
• You want to work in a repository with a terminal agent. See best AI coding tools.
• You need a store-ready mobile app and are comfortable with Android Studio: the Android output is native Kotlin, so you can carry it further.

If you are still looking for something to build, browse the ideas on the board. Each has a generated prompt you can adapt for Build mode.