Google AI Studio

The Build agent acts on outside services as well as code: asked to add a database, it provisions Firestore and Firebase Authentication with Sign in with Google and writes the integration code, and it can connect Supabase or MongoDB Atlas instead.

Twelve Workspace APIs, among them Gmail, Calendar, Drive, Docs, Sheets, Slides and Tasks, are wired from chat or the Integrations panel with the OAuth flow generated, so each end user authorizes their own data.

The agent installs npm packages on request, reads third party keys such as Stripe or SendGrid from server side secrets, and pushes changes to GitHub with commit messages it writes.

Sourceai.google.dev/gemini-api/docs/aistudio-fullstackread 2026-10-18

A single prompt produces a working project, a React front end with a Node.js server by default or a Kotlin and Jetpack Compose Android app, and the agent carries context across every prompt and file state, edits many files in one change, resolves dependencies and verifies its execution as it goes.

Work continues through chat, direct edits in the Code tab, or annotation mode, where the builder highlights part of the running interface and describes the change. When a feature needs a key or a new environment variable, the agent stops to request it in a chat card, then wires it into the server code.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18

The agent's grounding is the project itself: it maintains context of the entire codebase across prompts and file states, so later requests build on what is already there, and imported GitHub repositories become that context. AI Chips add Gemini capabilities such as image generation and Google Maps data to the app being built. There is no place to give the builder standing knowledge, such as documents, coding rules or a design system it should follow across projects, beyond what is written into each prompt.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18

The agent pauses for a person when it needs a secret, showing a chat card that asks for the key before wiring it in, and apps stay private until their owner shares them or puts them on Cloud Run. When GitHub and AI Studio diverge, a Resolve conflicts dialog shows the two versions side by side for the builder to choose. Changes otherwise apply directly to the project: there is no plan to approve before the agent edits code and no approval gate on provisioning Firebase services.

Sourceai.google.dev/gemini-api/docs/aistudio-fullstackread 2026-10-18

Access runs on Google accounts, and Workspace administrators can turn AI Studio on or off for groups or organizational units, with users under 18 on Education accounts blocked. Apps are private by default and can be shared to view and fork or with edit permission, and the Gemini API key and other secrets stay on the server, out of shared code. On the free tier, content is used to improve Google's products, while paid tier content is not. AI Studio carries no single sign on, roles, audit log or attestation of its own beyond Google's account and Workspace controls.

Sourceai.google.dev/gemini-api/docs/workspaceread 2026-10-18

What the agent changed shows up in the Code tab and the live preview, and two way GitHub sync records each push as a commit with an AI written message, giving an outside history of the agent's work, with a side by side diff when changes conflict. Within AI Studio there is no log of the agent's steps, tool calls or provisioning actions, and no audit trail of who prompted what in a shared app. Apps published to Cloud Run are monitored through Google Cloud's own logging in the linked project, outside AI Studio.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18

Each app persists on the AI Studio Apps page with its code, chat and settings, the agent keeps context across prompts within a project, and two way GitHub sync keeps a durable copy of the code with a commit history. Deleting an app removes it from the Apps page. Nothing carries preferences or lessons from one project to the next, and there is no memory a builder can review or edit apart from the project's code and conversation.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18

The builder runs only as a hosted service at aistudio.google.com. Publishing creates a Cloud Run service, either on a Starter Tier that needs no billing account but allows two services in a single region, or as a standard deployment in a Google Cloud project the builder links, with billing enabled and higher quotas. Apps can also leave AI Studio as a ZIP download for any host or through GitHub, and published apps take a custom subdomain under ai.studio. There is no self hosted builder and no region choice on the Starter Tier.

Sourceai.google.dev/gemini-api/docs/aistudio-deployingread 2026-10-18

The App Gallery is a visual library of working projects that can be previewed and copied to remix, and AI Chips drop ready features such as image generation or Google Maps data into a new app. I'm Feeling Lucky has Gemini propose a whole project to start from, and remixed or imported projects arrive with their code, ready for the agent to change. Starting points cover web apps and Android apps alike.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18

Work starts when a person types or speaks a prompt in Build mode, asks Gemini for an idea with I'm Feeling Lucky, copies an app from the App Gallery to remix, imports a project from GitHub or highlights part of the preview in annotation mode. Pulls from GitHub bring in outside changes for the agent to work on. Every change begins with someone in the AI Studio interface: nothing starts the agent on a schedule, from an issue, a webhook or a monitoring alert, and there is no API for handing it a task.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18

The Build agent's model is neither named nor selectable, and everything in AI Studio runs on Google's own Gemini family, with premium models such as Gemini 4 Argon unlocked by paid plans and no models from other makers on offer. Apps the agent builds call Gemini through an automatically configured server side key. There is no bring your own model, no other provider and no routing between makers for the agent itself, which makes Google the only model supplier for anything built here.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18

Projects extend outward through two way GitHub sync, ZIP export for local development, npm packages the agent installs, server side secrets for any third party API and OAuth callback URLs listed under Integrations for services beyond Workspace. Build mode itself has no API, SDK, CLI or MCP connection for driving the agent from outside or adding tools to it, so the extension points are in the generated app rather than in the builder.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18

The agent verifies its execution to cut down on hallucinated code, every change renders in a live preview, Android apps run in a browser emulator or install on a phone over ADB, and multiplayer apps can be tested by opening Build mode in several tabs, with a dev container simulating extra players. Build problems are fixed by prompting the agent and resharing. There are no automated tests, test runs before publishing or checks on the generated app beyond what a person sees in the preview.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18

The live preview and the Android emulator display the app the agent has built so a person can try it, and annotation mode lets that person point at part of the running interface to ask for a change. Both serve the builder checking the agent's own output. Build mode gives the agent no control of a browser, a desktop or another application's interface, and its reach into other systems runs through Firebase provisioning, Workspace APIs, npm packages and server code, so a task inside a third party web portal is out of scope.

Sourceai.google.dev/gemini-api/docs/aistudio-build-moderead 2026-10-18