Independent guide. Not affiliated with, endorsed by or sponsored by Google LLC.
Official Antigravity site Official documentationIndependent field guide
Google Antigravity moves the unit of work from the line you are typing to the task you described. This is a plain, sourced account of what it does, what it costs today, and where it will and will not help.
A sketch of the artifact flow described in Google's announcement — not a capture of the product.
How it is put together
Antigravity is not one feature. It is a set of surfaces the agent can reach and a set of documents it hands back. Each part changes where your review effort goes, so it is worth knowing which you are buying.
Tab completion, inline commands and an agent you talk to in a side panel. You stay in the file. This is the mode to start in, because you can see every change as it lands and back out of it in one keystroke.
Google's description is agent-first: surfaces embedded into the agent rather than an agent embedded in a surface. You spawn work into parallel workspaces, then observe rather than watch — the point is that you are not present for each step.
The agent does not stop at the file. It runs your commands in the terminal and actuates a browser to exercise what it changed. That last part is what separates this from a completion tool: the agent can check its own work end to end.
Task lists, implementation plans, walkthroughs, screenshots and browser recordings. You comment on text artifacts the way you would in a document, and select-and-comment on the visual ones; the agent takes the feedback mid-run.
Agents read from and write back to a shared store of snippets and task sequences that worked. In principle the fifth task in a codebase should go better than the first.
Before you install anything
These are the ones that decide whether the tool fits how you work. The product questions — models, price, privacy, languages — are answered in full on the FAQ page.
Not on the strength of a preview. Give it a task you already know how to do, in a worktree, and read the plan before the diff — that is the cheapest way to find out whether the review load suits you.
The failure mode people report is not bad code. It is a plausible walkthrough for a change that solved a slightly different problem than the one you described.
Limits refresh on a five-hour cycle and scale with how much work a request represents, not with how many prompts you sent. One large refactor can eat more of a window than a whole afternoon of small questions.
Plan around the cycle, not the prompt count — and expect that number to change, because a free preview's limits are the first thing to move when a preview ends.
It is enough to see that a path was exercised. It is not enough to see that the path is correct, that nothing else broke, or that the change is the one you wanted. Treat a recording the way you would treat a colleague's screen share: useful, not authoritative.
An editorial guide written by readers of Google's own documentation. It hosts no installer, takes no account, processes no payment and sets no cookie. Every factual claim here names its source and the date it was checked.
It is not Google, and it is not a substitute for Google's documentation. Where the two disagree, Google is right and we are out of date — tell us and we will fix it.
What you need
Taken from Google's getting-started documentation. Rows marked changes are the ones most likely to be out of date first — check them against the source link before you rely on them.
| Item | Current position | Source |
|---|---|---|
| macOS | 12 Monterey or newer, Apple Silicon or Intel | |
| Windows | 10 or newer, 64-bit — x64 or ARM64 | |
| Linux | glibc 2.28 or newer — Ubuntu 20, Debian 10, Fedora 36, RHEL 8 | |
| Where to get it | antigravity.google/download only. No third party redistributes it. changes | |
| Project setup | New project → add local folders or Git repositories → Create. That folder set is the boundary of what agents can touch. | |
| Run mode | Local Mode edits the folder directly; New Worktree Mode isolates the run in its own Git worktree. | |
| Cost | No charge during the public preview. | |
| Metering | Limits refresh on a five-hour cycle and scale with task complexity. changes | |
| Account and payment | Handled entirely by Google. This site never takes either. never here |
The evidence
The reference this site was modelled on carries six named developer testimonials. We have none, because we have no way to verify one. Sourced statements are the substitute.
Antigravity ships an Editor view — tab completion, inline commands and an agent in a side panel, for working alongside you — and a Manager view, which Google describes as mission control for spawning, orchestrating and observing several agents at once across parallel workspaces.
Source: Google, “Introducing Google Antigravity” · Checked 29 August 2026
The same agent operates across the editor, the terminal and a browser. Browser control is the part that separates it from a completion tool: the agent can drive the app it just changed and check the change landed.
Source: Google, “Introducing Google Antigravity” · Checked 29 August 2026
Work comes back as task lists, implementation plans, walkthroughs, screenshots and browser recordings. The stated intent is that you can verify what an agent did without reading every line it wrote.
Source: Google, “Introducing Google Antigravity” · Checked 29 August 2026
Text artifacts take Google-Docs-style comments; visual artifacts take select-and-comment feedback. The agent picks the comments up mid-run rather than at the end.
Source: Google, “Introducing Google Antigravity” · Checked 29 August 2026
Agents both read from and contribute to a shared knowledge store — code snippets and task sequences that worked — so later runs start from earlier ones.
Source: Google, “Introducing Google Antigravity” · Checked 29 August 2026
Google runs Antigravity as a public preview at no charge. Rate limits refresh every five hours and are tied to how much work a request represents, not to how many prompts you sent.
Source: Google, “Introducing Google Antigravity” · Checked 29 August 2026
Several write-ups place an Antigravity 2.0 at Google I/O in May 2026, with the desktop app rebuilt around multi-agent orchestration and a new command-line tool written in Go. We have not found this on a Google page, so treat the version number and the date as unconfirmed.
Coverage lists Gemini 3 Pro and Flash alongside Anthropic Claude and an open-weights GPT model, with the non-Google models sometimes requiring your own API key and billed separately. Google's own announcement names Gemini 3, Claude Sonnet and GPT models without pinning the commercial arrangement, so check the model picker in your own install before planning around it.
A figure of five simultaneous agents in Manager view appears in more than one review. It is plausible and it is not documented anywhere we can cite.
If you can point us at a Google page that confirms or contradicts any of the three above, send it over and this section changes the same day.
Where to go next
Downloads, accounts, billing and support are Google's, and only Google's. This guide exists to help you decide whether to spend an afternoon on it — not to stand between you and the product.
This site has no installer, no login, no cart and no cookie banner, because it has nothing to put in one.
Never enter a Google password, a card number or an API key on this site. Nothing here would legitimately ask for one.