Independent community project. Not affiliated with Google LLC. Read the disclaimer
Antigravity Community Get Antigravity
  1. Home
  2. Guides

Guides that assume you have work to do

Five short, practical sections. No screenshots of a menu you can find yourself — just the decisions that matter and the failures worth knowing about in advance.

Install it in about ten minutes

Antigravity ships as a normal desktop application. Download it from Google's official download page — not from a mirror, and not from a site that repackages it. Verify the domain reads antigravity.google before you run anything.

Windows

  1. Run the installer. Windows 10 64-bit or later is required.
  2. If SmartScreen warns you, confirm the publisher before choosing to continue.
  3. On first launch, sign in with a Google account. This is what unlocks model access.

macOS

  1. Open the disk image and drag the application to /Applications.
  2. macOS 12 Monterey or later is required. Both Apple silicon and Intel builds exist — take the one matching your machine.
  3. Grant screen-recording permission if you intend to use browser verification; the agent captures screenshots through it.

Linux

  1. Check your C library first: ldd --version must report glibc 2.28 or newer.
  2. Distributions older than roughly 2019 will not run it without upgrading.
  3. Install the package for your distribution, then launch it from your application menu so the desktop entry registers correctly.
Bringing your VS Code setup across Antigravity is built on the VS Code codebase, so the first-run wizard can import settings, keybindings and extensions. Import, then immediately test the two or three extensions you genuinely cannot work without — the ones that hook into VS Code internals are where compatibility breaks.

Your first agent run

Resist the urge to hand it a rewrite. The fastest way to calibrate your trust is a small task with a symptom you can see.

  1. Pick a visible bug. A button that does nothing, a layout that breaks at a specific width, a form that accepts something it should reject.
  2. Describe the symptom, not the fix. "The promo field accepts expired codes" beats "add a date check in validateCoupon". You want to see how it diagnoses.
  3. Read the plan before approving. This is the cheapest review you will ever do. If the plan names the wrong file, correct it there rather than after the diff.
  4. Let the browser step run. Point it at your dev server and let it reproduce the bug first, fix it second, and reproduce again.
  5. Read the walkthrough. Commands, diffs and screenshots in one place. If the evidence does not convince you, the change is not done.

A workflow that holds up

Two habits separate teams that get value from agents from teams that quietly stop using them.

One task, one workspace. The Agent Manager exists so that a long refactor and a quick fix do not share context. Give each its own workspace and the transcripts stay readable.

Correct on the artifact. When the plan is wrong, comment on the wrong sentence. When a screenshot shows the wrong state, comment on that region. Retyping the entire request throws away the context the agent already had.

Beyond that: keep task groups aligned to features rather than to files, write recurring conventions into the knowledge base once instead of repeating them, and treat the walkthrough as the artifact you attach to the pull request. A reviewer who can see the commands and screenshots does not have to rerun anything.

Working on code you cannot leak

Any hosted model means your code leaves the machine when you invoke it. That is not unique to Antigravity, and pretending otherwise helps nobody. What you can control is scope.

  • Check your organisation's policy before pointing an AI editor at a private repository. This is a governance question, not a tooling one.
  • Keep secrets out of the working tree. Environment files, tokens and key material should not be in a directory an agent can read.
  • Review the terminal permissions you grant. A scoped shell is only scoped if you keep it that way.
  • Read Google's own privacy documentation rather than a third-party summary — including ours. It is the only authoritative source on retention and training use.

When something goes wrong

SymptomUsual cause
Agent stops mid-taskPreview rate limit reached. Limits refresh on a rolling window; splitting the task into smaller ones helps.
Browser step never startsDev server not running, wrong port, or the browser extension has not been granted permission.
An extension misbehaves after importIt depends on VS Code internals. Disable it, confirm, and look for an alternative.
Screenshots are blank on macOSScreen-recording permission was not granted. Grant it in System Settings and restart the app.
Commands fail that work in your own terminalThe scoped shell does not inherit your full profile. Set what it needs explicitly.
Still stuck? We are a community site and cannot provide product support. For defects and account issues, use Google's official channels. If one of our guides is wrong or out of date, send us the details and we will fix it.

Last reviewed 24 August 2026.