Google AI Studio 101: Building With Gemini in the Browser

Google AI Studio is the browser-based playground and developer environment for Google’s Gemini family of models. You go to aistudio.google.com, sign in with a Google account, and you’re inside a multimodal prompt console with access to the full model lineup (Gemini 3, Nano Banana for images, Veo for video, Lyria for music, Imagen, Embeddings, the Live API, the Robotics models), plus a free API key, plus a Build mode that generates entire full-stack and native-Android applications from a natural-language prompt. The point of Google AI Studio is to compress the loop between "I have an idea" and "I have a working prototype using Gemini" from days into minutes.

This is the foundational 101 for Google AI Studio. We cover what AI Studio actually is and how it relates to the Gemini API, what Build mode does and why the Antigravity Agent powering it matters, the full-stack and Android capabilities that shipped in the 2026 update, how to deploy what you build (Cloud Run, GitHub, ZIP export), the model catalog you can prototype against, how the API-key story works (and where to be careful with it), and who AI Studio is for. This post is a tour, not a tutorial; deeper walkthroughs of specific Build mode workflows will follow as the platform continues to evolve.

What Google AI Studio actually is

At its core, Google AI Studio is two things bundled into one browser app. It is a prompt console where you can pick a Gemini model, set parameters (temperature, top-p, system instructions, tool selection), feed in text, images, video, audio, or documents, and see what the model returns. It is also the easiest way to get a Gemini API key: clicking through to aistudio.google.com/apikey gives you a key you can use immediately from your own code. The same surface that lets you experiment interactively gives you the credential you need to ship.

What separates 2026’s AI Studio from the original 2023-era "playground for an LLM" tool is the Build mode. Build mode lets you describe an application in natural language and watch the system generate a complete project: a React frontend by default, a Node.js server-side runtime, dependency management, secrets handling, sometimes Firebase integration, and a live preview. The intent shift is real. AI Studio used to be where you went to test a prompt; it is now also where you go to test an idea by building it.

Underneath both surfaces, AI Studio is a thin client over the Gemini API. Everything you can do in AI Studio’s prompt console you can do programmatically through the same API. The console is just a particularly polished interface to it, designed to remove the friction of getting your hands on the model.

Build mode: vibe coding with the Antigravity Agent

Build mode is where AI Studio’s product narrative shifted most visibly in 2026. Google calls the workflow "vibe coding": you describe what you want in plain language ("a markdown editor with live preview and saved snippets in Firebase"), and the system generates the code, runs it, and shows you the result in a live preview pane on the right. You iterate by chatting with the system, by clicking the Annotation mode button and highlighting parts of the UI to describe the change you want, or by editing the generated code directly in the Code tab.

The model doing the generation is not just "Gemini wrapped in a UI." It’s the Antigravity Agent, the same agent harness that powers Google Antigravity (the agent-first development platform Google shipped in November 2025 and expanded into a multi-surface platform with Antigravity 2.0 at I/O 2026). The Antigravity Agent maintains context across prompts and file states, manages dependencies across multiple files, and verifies its own code updates to reduce hallucinations. That last detail matters: when the agent makes a change, it doesn’t just emit a diff and hope; it runs the result to confirm it didn’t break the project.

The practical effect is that Build mode isn’t trying to generate a one-shot code dump from a prompt. It is running an actual agent loop over your project, the same loop developers use inside the Antigravity IDE, but exposed inside a hosted environment with no install step and no local toolchain to manage. If you’ve felt the difference between "AI completion in your editor" and "an agent that owns the loop," that difference is what AI Studio’s Build mode brings to a browser.

A few quality-of-life features that round out Build mode: the App Gallery is a library of remixable starter projects, useful when you want a project shape rather than a blank canvas. The "I’m Feeling Lucky" button has Gemini generate a project idea and a starting prompt for you. AI Chips let you add capabilities (image generation, Google Maps grounding, others) to your prompt with a click rather than having to remember the right wiring.

Full-stack capabilities

The version of Build mode that ships in 2026 supports actual full-stack apps, not just client-side prototypes. The pieces:

A server-side runtime based on Node.js. Code that you don’t want exposed in the client (anything touching API keys, database credentials, server-side business logic) lives here. The agent installs npm packages as needed (and you can request specific ones), so the modern web ecosystem is available.

Secrets management in the Settings menu. API keys and other sensitive values get stored once and made available to the server-side code, without ever appearing in the client bundle. This is the obvious right answer to client-side key exposure that the AI-app-generation category has been struggling with.

Multiplayer support through the server-side runtime, including the state and connection plumbing required for real-time collaboration features. Apps that need shared cursors, live updates, or synchronous interactions between users get the server-side scaffolding without the developer having to design it.

Firebase integration for persistent storage and authentication. The agent can automatically provision Firestore (database) and Firebase Authentication (sign-in flows, including "Sign in with Google"), then write the application code that uses them. For a category of small apps where Firestore plus Firebase Auth covers most of what you need, this is the path of least resistance from prompt to working app.

For deeper background on what this developer-tool category looks like more broadly, our coverage of Google Antigravity covers the IDE-first form factor of the same agent loop, and our Claude Code vs OpenAI Codex comparison covers two of the major CLI-first competitors.

Building Android apps in Google AI Studio

The other 2026 expansion is native Android. You describe the Android app you want and the Antigravity Agent generates a complete Kotlin and Jetpack Compose project, the way you’d build a modern Android app by hand. The output is real Android code targeting the modern toolchain (not a wrapped web view), with the agent handling the project structure, dependencies, and UI composition that an Android engineer would otherwise set up by hand.

This is interesting strategically. Native mobile development has been the most-resistant category to AI-assisted generation because the integration points (Activity lifecycles, Compose state, Gradle, the SDK in general) are dense and unforgiving compared to web. Putting that workflow inside an AI-studio-style "describe and ship" environment is a meaningful step, and it’s also the clearest signal that the Antigravity Agent’s "verified execution" model is doing real work, because Android compilation failures get caught and corrected mid-flight rather than handed back to the user as a stack trace.

If your work touches mobile, the practical implication is that prototyping an Android app no longer requires you to set up Android Studio, install the SDK, configure emulators, and learn Gradle before you can see whether your idea makes sense. You can sketch the app in AI Studio first and graduate to a local Android Studio project (or to Android Studio’s own AI integrations) when the prototype merits the deeper investment.

Models you can prototype against

AI Studio gives you access to the current Gemini model lineup with no extra setup. The relevant pieces, as of mid-2026:

For text and reasoning, the Gemini 3 family (Gemini 3 Pro for the heavier work, Gemini 3 Flash for fast iteration) plus the newer Gemini 3.5 Flash announced at I/O 2026 (see our Gemini 3.5 Flash for AI coding coverage for the benchmarks and the coding-agent angle). The model picker exposes Thinking variants for the cases where you want the model to deliberate longer.

For images, Nano Banana (Google’s headline image-generation model, sometimes referenced in docs by the playful internal name) and the Imagen family for higher-fidelity image work.

For video, Veo for generation, plus video-understanding capabilities for analyzing existing video content.

For audio, Lyria 3 for music generation and Lyria RealTime for streaming use cases, plus text-to-speech and audio-understanding models.

For embeddings and retrieval, the Gemini embedding models that pair with vector databases for RAG workflows.

For robotics, the dedicated Robotics models for the more specialized use case.

You can switch between these freely in the prompt console. Build mode picks the right model for the work it’s doing automatically, but you can override if you want a specific model behind your application.

Sharing, deploying, and exporting your work

Once you’ve built something in AI Studio, you have three paths out.

Deploy to Google Cloud Run with a single click. Your app gets a public URL, Cloud Run handles scaling, and the deployment includes an API proxy that keeps your Gemini API key on the server side rather than exposing it in the client. Cloud Run’s standard usage-based pricing applies.

Push to GitHub. AI Studio’s GitHub integration creates a repository and commits your latest changes, so you can continue development in your own toolchain (or in the Antigravity IDE) with the project history preserved.

Download as a ZIP if you want to take the project entirely off-platform. The trade-off here is the API-key story (covered below); apps deployed outside AI Studio need to handle key security themselves.

Sharing inside AI Studio is also a real workflow. By default an app is private, but you can share with specific users (read-only or edit access), and the AI Studio proxy ensures that shared apps use the recipient’s API key rather than yours, so you’re not paying for someone else’s usage.

API keys and security

AI Studio’s developer experience around API keys is intentionally permissive to keep the prototyping loop fast, with explicit warnings about where the easy path stops being safe.

Inside AI Studio’s hosted environment, generated apps can call the Gemini API from the client. This is normally bad practice (it exposes your API key in client-side code), but AI Studio runs a proxy that attaches the key without exposing it. Apps deployed to Cloud Run from AI Studio retain that proxy pattern: the deployed app uses your API key, and the key stays server-side. Shared apps similarly proxy through the recipient’s key.

The risk shifts when you download the ZIP and deploy elsewhere. In a standalone deployment without an AI Studio proxy in front, calling Gemini from client-side code means embedding a real key in code that any user can see. The Google AI Studio docs are explicit about this: don’t put a real API key in client-side code outside AI Studio’s hosted environment. Move the key-using logic to a server-side component before deploying externally. Google’s own guidance is to use a placeholder value (process.env.GEMINI_API_KEY) during AI Studio development that gets substituted by the proxy, then replace the calling pattern (not the key) when deploying elsewhere.

For any production application, the right pattern is the standard one for AI APIs generally: keep the key on the server, expose your own API to the client, and rate-limit and authenticate as appropriate. AI Studio’s Build mode is excellent for prototyping; the leap from prototype to production is the place to do the security work properly.

Where Google AI Studio fits in the broader Google AI stack

It’s helpful to position AI Studio against the other surfaces in Google’s AI developer story.

The Gemini API is the foundation. AI Studio is a UI over it; everything you can do in AI Studio’s console and Build mode is available programmatically through the API.

Vertex AI is the Google Cloud-hosted enterprise platform for the same models, with the enterprise-grade controls (private endpoints, IAM, audit logging, regional data residency) that production-grade enterprise deployments need. The migration path from AI Studio prototypes to Vertex AI deployments is documented; the same model calls work in both places.

The Gemini Enterprise Agent Platform is the agent-execution layer on the enterprise side, paired with Managed Agents in the Gemini API for callable agent workflows. Apps you build in AI Studio that need to graduate to production agentic workflows would land here.

Google Antigravity is the IDE-first form factor for the same agent loop AI Studio’s Build mode uses. If your work is heavy on local code rather than browser-hosted apps, Antigravity is the same platform with a desktop IDE in front of it.

Workspace access for AI Studio (the access path for Google Workspace users specifically) is documented separately and matters for organizations that need AI Studio under their Workspace governance rather than as personal Google accounts.

The right mental model is: AI Studio is the prototyping and developer-onboarding surface, and the rest of the stack is where production workloads, enterprise governance, agent loops, and full development environments live. AI Studio is designed to make the first ten minutes with Gemini productive; the other surfaces are designed for the rest of the project lifecycle.

Who Google AI Studio is for

For developers who want to evaluate Gemini against their use case, AI Studio is the lowest-friction starting point. Sign in, get an API key, start calling the model. Build mode is the right next step when you want a working demo rather than a prompt experiment.

For solo developers and small teams shipping AI features into their own products, AI Studio’s Build mode and one-click Cloud Run deployment cover the path from idea to shipped prototype faster than the traditional "set up a project, wire up auth, configure CI/CD, then start" sequence. The trade-off is the API-key and deployment-pattern caveats covered above.

For students, hobbyists, and people learning to build with AI, AI Studio is meaningfully better than competing options because the model access is generous, the Build mode is forgiving, and the App Gallery gives you a library of working examples to remix and reverse-engineer.

For larger teams and enterprise customers, AI Studio is the on-ramp to evaluation, not the production endpoint. The production path is Vertex AI, the Gemini Enterprise Agent Platform, or Antigravity (for developer tooling), each of which has the governance and scale story that AI Studio’s "personal Google account, browser-hosted, free tier" framing doesn’t try to provide.

Frequently Asked Questions