Questions, answered without marketing

Three concrete differences. It runs several agents at once across separate workspaces through the Agent Manager. It can drive a real browser, so an agent can verify its own change against the running application. And it produces artifacts — plans, screenshots, command logs — as reviewable evidence. Cursor and Windsurf both have strong agents; neither currently controls a browser.

Gemini 3 Pro from Google, Claude Sonnet 4.5 from Anthropic, and OpenAI's open-weight GPT-OSS models. You choose per task, which lets a cheaper model handle mechanical edits while a stronger one handles design decisions.

Windows 10 64-bit or later; macOS 12 Monterey or later; on Linux, glibc 2.28 or newer and glibcxx 3.4.25 or newer. Older long-term-support Linux distributions will not run it without upgrading the C library.

Generally yes. Antigravity is built on the VS Code codebase, so settings, keybindings and most Marketplace extensions import through the first-run wizard. Extensions that reach into VS Code internals are the common exception — test the ones you depend on straight after importing.

It is a dashboard for running several agents at the same time, each in its own workspace with its own task and progress trail. It replaces the single scrolling chat window with a queue you can scan, which matters once more than one job is in flight.

Antigravity can drive a Chrome session directly. After making a change, the agent starts your application, opens the relevant page, interacts with it, captures screenshots and reports what it observed. You can comment on any region of a screenshot and the agent treats that as its next instruction.

Code is transmitted for processing whenever you use an AI feature — that is inherent to any hosted model, not specific to this product. Google states that Antigravity code is not used to train its models and that the service follows Google's security standards. Treat Google's own privacy documentation as authoritative, and check it against your organisation's policy before using it on proprietary code.

The same range VS Code covers, because it is the same editor foundation: JavaScript and TypeScript, Python, Java, C#, C and C++, Go, Rust, PHP, Ruby, Kotlin, Swift and more. Agent quality varies by language in the way any model does — it is strongest where public training data is densest.

Google describes preview limits as generous and as refreshing on a rolling five-hour cycle. In practice most day-to-day use does not reach them; a long autonomous run across a large codebase can. Splitting big tasks into smaller ones is the usual workaround.

Not at the time of writing. MCP support has been signalled but had not shipped when this page was last reviewed. If your workflow depends on MCP today, Cursor and Windsurf already have it.

No. This is an independent community publication. We are not affiliated with, endorsed by, or sponsored by Google LLC, and we have no access to unreleased information. Everything here comes from public documentation and hands-on use. See our disclaimer.