How to Use Google Antigravity: Your First Agent Task, From Install to Review

CodexFuse Team · 2026-10-28T05:53:30.654Z · 15 min read

Google Antigravity is Google's agent-first dev platform. Install it, run your first agent task, pick the right model, and review the code before it ships — free.

Learning how to use Google Antigravity takes about an hour from download to your first finished agent task. Antigravity is Google's agent-first development platform: instead of autocompleting your code line by line, you describe a task in plain English and an AI agent plans it, writes it, tests it in a browser, and shows you the result for review.

It is free for individuals, runs on macOS, Windows, and Linux, and the whole beginner flow is install, first task, pick a model, review the work. If you already use Google AI Studio, you have met Antigravity's agent before — AI Studio's Build mode is powered by the Antigravity Agent — but this guide covers the full desktop platform, not the browser workspace.

What is Google Antigravity?

Google Antigravity is Google's agent-first development platform for building software with autonomous AI coding agents. You give an agent a goal — a feature, a bug, an investigation — and it reads your code, edits files, runs terminal commands, opens your site in Chrome to check the result, and hands you a plan, a diff, and a walkthrough to review.

The platform comes as four products on one agent harness: a desktop app that manages several agents at once, a terminal CLI, a full agentic IDE, and a Python SDK for building your own agents. Google lists the individual plan as available at no charge.

The shift is in who does the typing. A chatbot IDE answers one prompt at a time. Antigravity dispatches agents that work through a task end to end, including spinning up a browser to verify the interface actually works.

That is the idea to hold in your head for the rest of this guide: you are the reviewer now, not the typist.

Is Google Antigravity the falling Google logo trick?

No — and the name confusion is real enough to cost you an afternoon. Search results for this exact query still surface a guide to the old Google Gravity Easter egg, the browser toy where the Google homepage collapses under simulated physics. That experiment lives at mrdoob.com and has nothing to do with Google's developer platform.

The product in this guide lives at antigravity.google and is a coding platform, not a toy. If the page you landed on talks about dragging the Google logo around with your mouse, you are in the wrong place.

Close it and start at Google's official Antigravity site instead. Everything below — the install, the agent task, the models — is about the developer platform.

How do I install Google Antigravity?

Installation is a normal desktop download, not a waitlist or a cloud signup. Follow these five steps and you will be inside the app in minutes.

1. Check your machine. Antigravity runs on Windows, Mac, and Linux. Have at least 8 GB of RAM, with 16 GB recommended, because the app, your editor, and a dev server will all run at once.
2. Download it from antigravity.google. Use the official site only; the installer is the desktop app, and the same site offers the CLI and the IDE.
3. Sign in with a personal Google account. Antigravity is available for personal Google accounts in approved geographies, for users 18 and older. An @gmail.com address avoids most Workspace-account issues.
4. Pick review-driven development as your starting policy. This is the recommended starting point: the agent plans and works, but you stay in the loop and approve changes instead of the agent editing freely.
5. Open a folder or repository as your first Project. A Project groups the code, settings, and agent conversations for one task or product, and one Project can hold several related repositories.

That is the whole install. There is no separate signup form, no credit card, and no key to paste in — the free plan starts the moment you sign in. If your work or school account is blocked, that is an admin restriction on that account, not a problem with the installer; switch to a personal account.

Your first agent task, step by step

Pick a small, real task for your first run — not your hardest bug. The classic first task is a to-do app, and it is still the best choice: it is small enough to finish, big enough to show you the whole loop, and easy to judge when it is done.

1. Start a new agent task. In the desktop app, open your Project and start a conversation with an agent. In the terminal, the same flow starts with the Antigravity CLI command agy.
2. Choose Planning mode. Planning mode breaks the task into a structured plan with detailed artifacts before writing any code. Fast mode skips planning and executes immediately, which suits quick fixes — but for a first task, the plan is the lesson.
3. Choose New Worktree mode for anything risky. Local mode works directly in your existing checkout. New Worktree mode opens an independent Git worktree, so the agent's changes stay isolated until you approve them.
4. Write the task in four parts. Use this template for every first task, and most later ones too: the context (what the code is), the action (what to build or fix), the constraint (what not to touch), and the check (how you will know it worked).
5. Read the plan before the code. The agent produces an implementation plan first. Comment on it like a document — this is where you catch a wrong approach cheaply, before any file changes.
6. Review the artifacts, then the diff. The agent shows a task list, code diffs, screenshots, and a browser recording or walkthrough. Read the diff, run the app yourself, and only then accept the changes into your branch.

Here is the four-part prompt shape, filled in with the example Google itself uses for a first task in an existing web app:

The check matters more than the rest. An agent that can open your page in Chrome and record what happened is an agent you can trust — and an agent whose claim you can verify in thirty seconds. If your task has no check, you are not delegating; you are hoping.

Planning mode vs Fast mode: which one should you pick?

Planning mode is the right default for anything you have not done before. It produces more artifacts and takes longer, but the plan is where you steer. Fast mode is for work you could already do yourself in five minutes.

The rule is simple: if the task touches more than one file, or you would want a colleague to explain their approach before starting, use Planning mode. If you already know the exact change and just want it typed, use Fast mode. Beginners should stay in Planning mode for their first week — the extra minutes buy the understanding that makes the next task faster.

Which model should you pick in Antigravity?

Start with Gemini 3.8 Flash. It is the default model, it is available on every plan including free, and it handles everyday agent tasks — the first to-do app, the settings page, the failing test — without burning your weekly quota on a model you do not need yet.

Two dates matter here. Gemini 4 Argon arrived in Antigravity on October 13, 2026, for Google AI Ultra and eligible enterprise customers — it is the model for tracing a failure through a large codebase, investigating logs and commit history, and planning changes across several modules.

The Claude 4.6 models leave the free column on November 2, 2026, after which the free tier holds Gemini 3.8 Flash alone. If you want to try a Claude model free, do it before that date, and keep your routine work on Flash so the quota lasts.

One more limit: Argon works in the desktop app, the CLI, and the IDE extensions, but not in the original Antigravity IDE itself. If Argon is the reason you are paying for Ultra, use the desktop app or the CLI.

What does the free plan actually cover?

The free plan is a real plan, not a demo. Individuals pay $0 a month and get the desktop app, the CLI, scheduled tasks, Gemini 3.8 Flash, unlimited Tab completions, and access to every product feature. The catch is the quota: free accounts get a weekly agent allowance that refreshes once a week, and Google publishes no number for it.

That quota tracks agent work, not prompts. A quick edit costs little; a long multi-file refactor can use a large share of the week.

Two habits make it last: keep tasks small and well specified, and save the Claude models for the tasks that need them before November 2. Your remaining usage per model sits on the Settings page, and the CLI shows it in the status line.

What the free plan does not give you: overage. Only Pro and Ultra subscribers can buy AI credits past the baseline, and the setting that controls it — AI Credit Overages — should stay on Never until you know your usage.

There is also no bring-your-own-key option: you cannot paste a Gemini API key into the app to lift the free cap. When the weekly quota is gone on free, you wait for the refresh — or upgrade to Pro, where the allowance refreshes every five hours until the weekly limit is reached.

The four ways to use Antigravity: App, CLI, IDE, and SDK

Antigravity 2.0, released in May 2026, turned the original AI coding editor into a family of products on one agent harness. You do not need all four — most beginners need one — but knowing the map stops you from fighting the wrong surface.

The desktop app is the main workspace: it groups conversations into Projects, runs several agents in parallel across workspaces, and automates routine work with scheduled messages. The CLI brings the same agent harness to the terminal, with conversation history, slash commands, subagents, and headless mode for scripts and automation jobs — and in June 2026 it replaced Gemini CLI for free and individual Pro and Ultra users.

Editor extensions bring agent assistance into VS Code, Visual Studio, JetBrains IDEs, Zed, and Xcode for developers who prefer their own editor. The SDK is a Python framework for building agents on the harness, including local-model setups — it can run agents on local models without an API key, with a recommended 24 GB of VRAM or unified memory for the Gemma 4 26B model.

Two features beginners underuse: scheduled tasks, which run an agent prompt on a recurring schedule — checking a repository for outdated dependencies, or collecting a daily summary of test failures — and Remote Control, which lets you check a running agent session from a browser on another computer, tablet, or phone while the session keeps running on your machine.

5 mistakes beginners make with Google Antigravity

Most first-week frustration with Antigravity is not the model. It is the workflow around it. Avoid these five and the tool feels dramatically better.

1. Starting without Git. The agent edits real files. If you skip version control, a bad agent run can lose work you cannot get back — commit first, or use New Worktree mode so the agent works in an isolated copy.

2. Writing vague prompts. "Fix the login" is not a task. Name the files, the expected behavior, and how you will check the result — the four-part shape from the first-task section — because the clearer the intent, the less the agent guesses.

3. Skipping the review. The plan, the diff, and the walkthrough exist to be read. Never merge a diff you have not read — review-driven development is the recommended starting policy for a reason.

4. Routing your login through another tool. Using third-party software to access Antigravity violates Google's Terms of Service and can suspend your account. Headless agy -p is for your own scripts and CI workflows — use Antigravity directly.

5. Burning the weekly quota on one giant task. Quota tracks how much work the agent does, so one sprawling "rebuild my app" prompt can eat the week. Break the work into small tasks, run them one at a time, and keep the hard-model spend for the jobs that earn it.

Google Antigravity vs Cursor vs Claude Code

Antigravity, Cursor, and Claude Code all put an AI agent on your code, but they start from different places. The right pick depends on where you already work.

The clearest separation is parallel work: Antigravity runs multiple agents across different workspaces at once — one refactoring the data layer while another writes tests — while Cursor and Copilot-style tools operate synchronously. If you want to direct a small team of agents and review what each one did, Antigravity is built for that.

If you want one agent beside your cursor in the editor you already live in, Cursor fits better. And if you want the agent in the terminal, Claude Code is the equivalent — which is also why Google Gemini matters here: the same Gemini models power the consumer chatbot, AI Studio, and Antigravity, so the model you already know from the Gemini app is the one running your first agent task.

Sources

• Google Antigravity — official homepage (antigravity.google) — the four surfaces (App, CLI, IDE, SDK) and "Available at no charge". Read October 28, 2026.
• ScriptByAI, "Google Antigravity 2: Agent-First Dev Platform with Gemini 4 Argon" (scriptbyai.com/google-antigravity-dev-platform) — Antigravity 2.0, the model-availability table, the free tier and weekly quota, CLI agy, worktree mode, scheduled tasks, Remote Control, Argon rollout October 13, 2026, and the June 2026 Gemini CLI transition. Published October 14, 2026. Read October 28, 2026.
• Sharkly, "Is Google Antigravity Free? What the Free Plan Covers (2026)" (sharkly.ai/blog/is-antigravity-free) — free plan $0 a month, weekly quota with no published number, Claude 4.6 models removed November 2, 2026, personal-account sign-in rules, no bring-your-own-key, no overage on free, and the Terms of Service note on third-party wrapping. Read October 28, 2026.
• OpenReplay Blog, "A Beginner's Guide to Google's Antigravity IDE" (blog.openreplay.com/google-antigravity-ide-guide) — Planning vs Fast mode, Agent Manager vs Editor, artifacts, and review-driven development as the recommended starting policy. Published March 29, 2026. Read October 28, 2026.
• Julian Goldie, "Google Anti-Gravity IDE: The Agent-First Coding Revolution Starts Now" (juliangoldie.com/google-anti-gravity-ide) — 8 GB RAM minimum and 16 GB recommended, the to-do app first task, multiple agents at once, and the Git and review warnings. Read October 28, 2026.
• CodexFuse, "How to Use Google AI Studio: The Beginner's Guide" (codexfuse.dev/blog/how-to-use-google-ai-studio) — AI Studio's Build mode is powered by the Antigravity Agent. Read October 28, 2026.

Frequently Asked Questions

Is Google Antigravity free?

Yes. The individual plan is $0 a month and generally available, with a weekly agent quota, unlimited Tab completions, and access to every product feature, including the CLI and scheduled tasks. The free model list shrinks on November 2, 2026, when the Claude 4.6 models are removed and Gemini 3.8 Flash remains the free Gemini model.

What is the difference between Google Antigravity and the Google Gravity Easter egg?

They share a name and nothing else. The Easter egg is an old browser toy where the Google homepage falls apart under simulated gravity. Antigravity is Google's agent-first development platform at antigravity.google — a desktop app, CLI, IDE, and SDK where AI agents write, test, and verify real code.

Do I need to know how to code to use Google Antigravity?

Not to start: you describe tasks in plain English and the agent does the typing. But you still need to review and verify the agent's work, so basic coding knowledge helps enormously — you are the reviewer, and the quality of the review decides the quality of the result.

Which model should beginners start with in Antigravity?

Gemini 3.8 Flash. It is the default model, free on every plan, and it handles everyday agent tasks without burning your weekly quota. Save the Claude models for harder jobs before November 2, 2026, and leave Gemini 4 Argon — Ultra and enterprise only — for multi-step work across large codebases.

Can I use Antigravity inside VS Code instead of its own IDE?

Yes. Antigravity extensions bring agent assistance to VS Code, Visual Studio, JetBrains IDEs, Zed, and Xcode. Model availability differs across these products — Gemini 4 Argon, for example, works in the app, the CLI, and the extensions, but not in the original Antigravity IDE itself.