Google AI Studio Antigravity: Full-Stack Apps in One Prompt

When Google launched a major upgrade to AI Studio on March 18, 2026, it wasn't just another chat interface refresh. The company embedded an autonomous coding agent called Antigravity directly into the browser-based build environment, then wired it to Firebase so that data storage, authentication, and hosting get provisioned the moment a prompt signals the need for them.

The result is a full-stack app builder that works entirely in a browser tab. No local environment, no config files. You describe what you want, approve the Firebase setup card when the agent asks for it, then publish to a live URL.

The interesting part is not the demo. It is the set of quotas, defaults, and one-way choices that sit underneath the generated app, because those are what decide whether the thing you built in twenty minutes survives contact with real users. This guide covers what the agent provisions, the exact free-tier ceilings Google publishes for it, and where the tool stops. If you are weighing a desktop AI IDE instead, see our Cursor vs Windsurf vs Zed comparison, or run your requirements through the vibe coding tool picker.

What Is the Antigravity Agent#

Antigravity is not a code completion overlay. It is an agent that plans, writes, tests, and debugs multi-step applications in response to natural language prompts. Google launched it as a standalone IDE in November 2025, then wired the same agent into AI Studio's browser-based Build mode in March 2026. Two surfaces, one agent. That distinction matters later, because the plans and model menu documented for the standalone IDE are not identical to what Build mode exposes.

The lineage runs back to Google's July 2025 deal with Windsurf. Google paid roughly $2.4 billion for a nonexclusive license to Windsurf's technology and hired its CEO plus a core group of engineers into DeepMind. It did not buy the company. Cognition acquired what remained.

On models, the standalone IDE's model documentation lists a selector rather than a single fixed engine: Gemini 3.7 Flash, 3.6 Flash, 3.5 Flash, and 3.1 Pro, alongside Claude Sonnet 4.6 (thinking), Claude Opus 4.6 (thinking), and GPT-OSS-120b. Image generation goes to Nano Banana 2 and is not user-selectable. Google does not publish a per-surface default, so treat "which model just wrote this" as something you set deliberately at the top of a session rather than something you can assume.

What makes Antigravity distinct from most vibe coding tools is backend-awareness. The agent reads semantic intent from your prompt. If you describe a task manager, it infers the need for persistent storage. If you mention "user login," it knows authentication is required. Rather than generating frontend-only code and leaving you to wire up a backend separately, Antigravity surfaces a prompt: "Enable Firebase?" When you approve, the provisioning happens automatically.

How the Firebase Integration Works#

This is the part worth reading twice, because it changes what "done" means when an app is generated.

Describe a feature that needs storage or sign-in and the agent surfaces a setup card. Firebase's integration documentation adds a detail the launch post skips: the detection is not reliable on its own. "For best results, explicitly ask the agent to use Firebase as the backend for your app." Say the word Firebase in your prompt.

On approval, three things happen:

1. Cloud Firestore is provisioned as the database layer.
2. Firebase Authentication is enabled with Google Sign-In configured.
3. The agent writes the wiring into your codebase — a /src/lib/firebase.ts client and a firestore.rules file.

Publishing is where the common description of this tool gets it wrong. Share > Publish deploys to Cloud Run, not Firebase Hosting, in the same project. Build mode's own documentation confirms the shape of the generated stack: a React frontend by default, plus "a Node.js runtime that allows for secure API calls, database connections, and npm package usage." Your Gemini key is injected as a server-side secret and, per Google, is "never included in client-side code." Cloud Run pricing applies once usage grows.

The free backend quota, in actual numbers#

Every Firestore database created through AI Studio lands in a shared quota group. These are the published daily ceilings, and they are the honest answer to "how far does the free tier go."

Source: Firebase, Add Cloud Firestore and Authentication to your Google AI Studio app.

Two consequences follow, and neither is obvious from the demo. First, the group is shared, so one runaway prototype starves the others: "If you hit a daily quota limit, the service will pause for all databases within the group and resume the following day at around midnight Pacific Time." A pause, not a bill. Second, escaping the ceiling is manual. You attach a billing account, then click Upgrade database in the Firebase console for that specific database, after which all usage bills at standard Cloud Firestore rates.

One more thing to get right the first time. The Google Cloud project and region you pick for your first integrated feature apply to every Firebase and Google Cloud service you add afterward, and that choice is permanent. Pick the region your users are actually in.

Also worth knowing before you start: the Firebase integration is web-only. Build mode can generate an Android app as a Kotlin and Jetpack Compose project, but Google states the Firebase integration "isn't supported for Android apps at this time."

Real-Time Apps: What's Actually Different#

Antigravity is often described as the only vibe coding tool that does multiplayer. That framing does not survive a look at the competition. Lovable ships a Supabase integration whose realtime channels cover live shared state perfectly well, and collaborative editing exists on both platforms.

The real difference is narrower, and more useful to know. Firestore's live listeners come wired into the generated app without a second integration step, so a prompt like "build a collaborative whiteboard where multiple users draw at once" produces a data model that already syncs. You test it by opening the published URL in two tabs. That saves an integration decision, not an impossibility.

Then read the quota table again. Real-time updates draw on a 50,000-per-day shared budget, and a live listener fires per client per change. A whiteboard with five people in it consumes that budget far faster than a CRUD dashboard with fifty. For a demo or a hackathon judge, comfortable. For an always-on internal tool with a live view left open on someone's second monitor, do the arithmetic first.

Supported Frameworks and Project Types#

Google's launch post names three frontend frameworks for Build mode: React, Angular, and Next.js. React is the documented default, and it is the path most likely to be well-trodden.

Next.js gives you server-side rendering if the app needs to be indexable. Angular is there if your team already lives in TypeScript and Angular conventions. All three come out with the Firebase client generated and pointed at your provisioned Firestore instance, so there is no manual initializeApp() step after generation.

If your stack is Vue, Svelte, or anything else, this is not your tool. That is a genuine constraint rather than a rounding error, and it is the first thing to check before you invest a session here.

Pricing: What the Free Tier Actually Buys You#

Getting started costs nothing, and the agent quota is the meter that eventually bites. It ties to Google's AI subscription tiers rather than to AI Studio itself.

The first thing to understand is that Google no longer counts requests. Antigravity's plans documentation states that "rate limits are correlated with the amount of work done by the agent," so a one-line CSS fix and a four-file refactor draw down very different amounts. Any "N prompts per day" figure you read, including older ones from Google itself, describes a scheme that has since changed shape.

Source: Google Antigravity Docs, Plans. Google publishes no exact numeric limits for any tier.

The practical read: on the free tier a heavy Saturday can cost you the rest of the week, since the refresh is weekly rather than daily. Paid tiers recover every five hours, which is the difference between a bad afternoon and a stalled project. Credit overage is opt-in through an "AI Credit Overages" setting with a Never or Always toggle, so an unattended agent cannot quietly spend on your behalf unless you let it. There is no bring-your-own-key path and no organizational contract tier for raised limits yet.

Subscription pricing, from Google's own pages: AI Plus is $9.99/month and AI Pro is $19.99/month. At I/O 2026 Google introduced a $100 AI Ultra plan and cut the top Ultra tier from $250 to $200. Regional Google One pages have lagged those changes, so check the price shown at checkout in your own country before budgeting.

How to Start a Project#

The workflow requires a Google account. No installations.

Step 1: Open AI Studio Build Mode

Navigate to aistudio.google.com. Select the Build tab in the left sidebar. This opens the Antigravity-powered coding environment.

Step 2: Write a Prompt

Describe the app in plain English, and name Firebase explicitly. Google's own guidance is that intent detection alone is unreliable, so the word does real work here:

Build a task manager where users can sign up with email,
create personal task lists, mark tasks complete, and see
task counts on a dashboard. Use Firebase as the backend.

Step 3: Approve Firebase Provisioning

The agent pauses and shows a setup card. Before you click through it, open the settings icon and check the Google Cloud project and region, because that pair is locked in for every service you add later. On approval the agent provisions Firestore, enables Authentication with Google Sign-In, and generates /src/lib/firebase.ts and firestore.rules in the project.

Step 4: Iterate in Natural Language

The built-in preview shows the running app. Feedback in the chat updates the code:

Add a due date field to each task. Show overdue tasks in red.

Step 5: Publish

Use Share > Publish. The app deploys to Cloud Run in the same project and you get a URL you can send to someone. Two caveats. Cloud Run pricing applies as usage grows, and if you share an app that calls the Gemini API, those calls count against your own usage limits rather than the visitor's.

If you would rather leave, export the project as a ZIP or push it to a GitHub repository. Running it elsewhere means setting GEMINI_API_KEY yourself in whatever hosting environment you move to.

Developer Considerations#

When it makes sense:

You want a deployed full-stack prototype inside a single session rather than an afternoon of wiring. Firebase is already an acceptable backend for you, so the provisioning removes work you would have done anyway. The app needs live shared state. React, Next.js, or Angular is a fine output.

When it doesn't:

You need a database that isn't Firestore. Moving to PostgreSQL, PlanetScale, or your own API means keeping the generated frontend and rebuilding everything behind it. You want line-by-line control, which agent-written code does not give you: iterations introduce regressions, and each fix is another prompt. Your frontend is Vue or Svelte. Or your usage pattern outgrows a weekly-refreshed free quota, which happens faster than most people expect.

Firebase lock-in:

The generated app depends on a Firestore data model the agent chose. Migrating off Firebase later means transforming both that model and every SDK call built on it. Real weight for a production service, near-zero for a prototype you would happily rewrite.

When to Use Antigravity, and When to Skip It#

The trade-off is speed against control. Run down this checklist before you build anything you plan to keep.

Reach for Antigravity when most of these are true:

• You want a working, deployed full-stack app inside a single session, not an afternoon of wiring.
• Firestore, Firebase Auth, and a Cloud Run deployment are all acceptable to you.
• Your frontend is React, Next.js, or Angular, the three frameworks Build mode supports.
• Projected traffic fits inside 40K writes, 50K reads, and 50K real-time updates a day, shared across every AI Studio database you own.
• You know which Google Cloud project and region the app should live in, because that choice is permanent.
• The project is a prototype, internal tool, hackathon build, or PoC rather than a long-lived production service.

Skip it, or plan to outgrow it, when any of these hold:

• You need a non-Firebase database (PostgreSQL, PlanetScale, a custom API). Migration means rewriting both the data model and the SDK calls.
• You want line-by-line control of the code the way Cursor or Claude Code give you.
• The service has a long maintenance horizon where agent-driven regressions become expensive.
• The app is Android-first, since the Firebase integration covers web output only.
• You cannot afford to lose agent access until the following week if a heavy session drains the free quota.

A rough rule: if the app is something you would happily rebuild from scratch later, Antigravity is a strong launchpad. If it is something you expect to maintain and extend for years, treat the generated project as a starting point you will move off Firebase, not a foundation.

How It Compares to the Alternatives#

vs. Bolt.new: Both generate a working frontend quickly. The split is what happens when the app needs to remember something: AI Studio provisions Firestore and Auth from inside the session, while Bolt leaves the backend choice, and the integration work, to you.

vs. Lovable: Lovable is aimed at founders who want a polished MVP, and its Supabase integration gives it a real database and realtime channels, so the backend gap is smaller than it is often described. Its free plan grants 5 build credits a day, capped at 30 a month, which is a per-day rhythm rather than Google's weekly one. Choose on backend preference: Postgres and Supabase, or Firestore and Google Cloud.

vs. Cursor / Claude Code: Different category. Cursor and Claude Code are for developers working inside an existing codebase they intend to keep. AI Studio Build is for getting from nothing to a deployed URL. The audiences overlap; the jobs don't. See where each lands in our best AI coding agents comparison.

Practical Use Cases#

Internal tools: HR portals, inventory dashboards, internal admin panels. These apps follow predictable patterns (login, CRUD, dashboard) that Antigravity handles well, and they don't need custom infrastructure.

Game prototypes: Turn-based games map cleanly onto Firestore documents with live listeners, and the state layer arrives already wired. Watch the 50K daily real-time update budget once more than a few players are connected.

Hackathon submissions: Zero to deployed in hours, with a Cloud Run URL live before the demo slot.

Proof-of-concept validation: A shareable URL is enough to test whether a feature resonates before anyone commits to a production build. Remember that Gemini API calls made by visitors draw on your quota, not theirs.

FAQ#

Q: Does it work without Firebase?#

Yes, for frontend-only apps. The Firebase setup card only appears when your app needs storage or sign-in, and Google recommends asking for Firebase by name when you do want it.

Q: Can I export the code and run it locally?#

Yes. Download the project as a ZIP or push it to GitHub, then run it like any React, Next.js, or Angular app. You will need to supply GEMINI_API_KEY in the new environment, plus the Firebase config values from your project settings.

Q: Which model writes the code?#

There is a selector rather than one fixed model. Antigravity's documentation lists Gemini 3.7 Flash, 3.6 Flash, 3.5 Flash, and 3.1 Pro, plus Claude Sonnet 4.6, Claude Opus 4.6, and GPT-OSS-120b for subscribers with third-party model access. Pick before you start; the choice holds for the rest of the task.

Q: Is it available outside the US?#

Google AI Studio is available in most regions where Gemini API access is permitted, and Firebase provisioning follows Firebase's own regional support. Check the current region list at ai.google.dev/available_regions before planning around it.

Q: What happened to Firebase Studio?#

It is being retired, not absorbed. Google announced the sunset on March 19, 2026, disabled new workspace creation and signups on June 22, 2026, and will delete remaining data after March 22, 2027. Migration goes to either Google AI Studio or the standalone Antigravity IDE, and Google frames these as two separate destinations rather than one merged product. Core Firebase services such as Firestore, Authentication, and App Hosting are unaffected by the sunset.

What Effloow Added#

Google's launch post sells the magic moment: prompt in, deployed app out. Anyone deciding whether to build on it needs the ceilings, not the demo. Those are what we assembled here.

• The shared Firestore quota table, pulled from Firebase's integration documentation: 1 GiB of storage, 40K writes, 50K reads, and 50K real-time updates a day, shared across every database AI Studio creates for you, with a service pause rather than a bill when you cross the line.
• The plan structure as Google actually documents it, including the shift away from counting requests toward measuring agent work, and the weekly-versus-five-hour refresh split that decides how badly a heavy session hurts.
• Corrections to three things widely repeated about this tool: publishing goes to Cloud Run rather than Firebase Hosting, real-time state is not a capability competitors lack, and Firebase Studio is being retired rather than folded into Antigravity.
• A decision checklist built from the constraints that are irreversible or hard to unwind: the permanent project and region choice, the Firestore data model you would have to transform to leave, and the three supported frameworks.

Key Takeaways#

AI Studio's Build mode covers the full stack from one prompt: frontend, Firestore, Authentication, and a Cloud Run deployment, with no configuration files touched. That much lives up to the pitch, and for a prototype it is genuinely fast.

The constraints are quieter but they decide more. A shared free Firestore quota that pauses rather than bills. A free agent quota that refreshes weekly, so one long Saturday can cost you the week. A project and region locked at first setup. A generated data model you would have to transform on the way out. For hackathons, internal tools, and PoCs those are cheap prices. For a service you expect to run for years, treat what comes out of the session as a starting point rather than a foundation.