Antigravity

The choice rests on two things: how much control you need and where you sign in from.

Who it is for. Anyone who wants to try the agentic style of work for free: you set a task in words, the agent writes the code and brings back a report with a browser recording. The Gemini 3.8 Flash, Gemini 3.1 Pro and Claude 4.6 models are already inside, no keys of your own needed. It also suits people running several tasks at once, since the Agent Manager panel drives agents in parallel.

Who it is not for. Anyone who needs predictable limits. Google trimmed quotas four times over the winter without warning, and the practitioners' protest reached the press.

Need to work without the Google cloud. Antigravity is closed: its source code is not published, there is no local version, and interactions improve Google's models by default. Open alternatives Cline and OpenCode work with their own keys and local models.

If it works for you, section 3 shows installation. The choice against Cursor is covered in section 7.

What vibecoding isStarting from scratch: how to set tasks for agents with no coding experience

The core idea: the agent here is not a chat in the sidebar, but a worker with its own workspace.

Agent Manager. A separate panel where you assign tasks and watch several agents at once. Each agent has its own workspace, with inboxes collected into one list: in the official demo, three agents simultaneously handle spaceflight booking, exactly the scene in the hero video above.

Evidence-based work. The agent does not just deliver code: the task includes a report artifact with a plan, list of changes, tests, and a browser recording. The agent opens the browser itself, clicks through the built app, and shows that it works.

Editor inside. It is still a familiar environment based on VS Code: tabs, terminal, Tab autocomplete, and inline commands. Those used to a regular editor can work as before and add agents as they gain confidence.

Control from your phone. Since version 2.9.1 (Aug 20, 2026), agent sessions on your machine are visible from any browser: Remote Control sends a notification when an agent has finished or is waiting for an answer.

A plan before the code. Since version 2.17 (Sep 22, 2026) the /plan command makes the agent lay out a plan first. You read and edit it, and only then does the agent get to the code.

The product's home is a standalone app for macOS, Windows and Linux. Since August 2026 the agent can also be called from other editors.

How to install

Step 1. Download the app from the page antigravity.google/download for your system.

Step 2. Sign in with your Google account.

Step 3. Create a project, add a folder with code using the Add Folder button, and state the first task in words. The agent will offer two modes: Local Mode edits files directly in the folder, New Worktree Mode works in a separate copy.

Other editor. There are extensions for VS Code, JetBrains, Xcode and two more editors, with sign-in through the same Google account. For terminal fans there is a CLI: on Jun 18, 2026 it replaced Gemini CLI, which Google wound down for personal accounts. For agent developers there is an SDK.

Steps verified Sep 26, 2026 against antigravity.google/docs/getting-started

In benchmarks Antigravity is still at the bottom of the tables: it solves tasks less often than its neighbors.

At Artificial Analysis. An independent lab measured the Antigravity SDK, the same agent harness in library form, with the Gemini 3.8 Flash model. The pair scored 41.9 against 66 for the leader, Claude Code with Opus 5.5 (coding agents index, Sep 26, 2026).

On the board. In the agents ranking below, Antigravity's spot is set by demand, index_. The measurement columns in its row repeat the numbers above.

Antigravity has no separate price tag: you pay for a Google AI subscription, as with Gemini.

Free entry. The $0 plan came out of preview on Feb 26, 2026: unlimited autocompletions and an agent-work limit that refreshes once a week. The limit is spent by the volume of the agent's work, so more simple tasks fit than complex ones.

Paid tiers. The Google AI Pro subscription costs $19.99 a month, Ultra comes in two flavors: $100 for five times the limits and $200 for twenty times. On paid plans the limit refreshes every five hours. Above the limits you buy AI credits.

Quotas are being cut. From December 2025 to March 2026 Google quietly trimmed limits four times, with no email and no line in the changelog. Practitioners did the math: the $20 plan came to give four times more than the free one instead of the earlier thirty-three times. The protest reached the press on Mar 12, 2026.

We look at three things: where your data goes, what the agent is allowed to do and what you are forbidden to do.

Data trains Google by default. The terms of service say outright that interactions are used to improve Google's products and models. It can be switched off in settings, and deleting history goes through an email to support.

Autonomy needs a leash. In December 2025 an agent in fast Turbo mode wiped a user's entire drive: the delete command went through without confirmation and past the trash. Under Google's terms, you are responsible for the agent's actions. Give it the first tasks in a test folder and don't turn on the no-confirmation mode on an important project.

Other pages can command the agent. A week after launch, PromptArmor researchers showed that an instruction hidden in small type on a web page can make the agent leak data externally. Google initially called it intended behavior; no confirmed fix as of August 2026. Protection is the same as for any agent with a browser: do not keep unnecessary secrets in the project and read reports before accepting work.

Third-party software is banned. You can't hook the Antigravity quota up to OpenCode or Claude Code. The documentation calls this a violation of the terms, and Google bans accounts for it.

The split is genre-defining, and Habr put it most precisely: Cursor remains an editor with a strong assistant inside; Antigravity builds an environment where agents carry out tasks.

Different center of gravity. Cursor was built around daily code writing: strong autocompletions, familiar editor, maturity. Antigravity targets a different scenario: you assign tasks to a fleet of agents and accept work from reports with evidence.

Different cost of risk. Both have a $0 tier, then Cursor costs from $20 per month, while Antigravity hides its price in a Google AI subscription from $19.99. The difference is predictability: it is Antigravity that carries the history of quiet quota cuts.

Who else is nearby. Claude Code lives in the terminal and wins on depth with Claude models. Jules, Google's second agent, runs in the cloud with no install.

Open Cline and OpenCode eliminate the geography issue entirely: your own keys, any models, local execution.

The product is ten months old: a loud launch, a winter of quota and security scandals, a summer turn toward enterprise subscriptions and new Gemini models every few weeks.