Google Antigravity review: what the App, CLI, IDE extensions and SDK do

Google Antigravity review in one line: Google’s agentic development platform has four surfaces that share one agent, a desktop App for running several agents in parallel, the agy CLI, extensions for six editors plus the Antigravity IDE, and a Python SDK, with a $0 Free plan and Google AI Plus, Pro or Ultra for higher limits, according to Google’s Antigravity site and docs overview.

Applies to: Google Antigravity 2.5+ (App, CLI, IDE extensions, SDK), per Google’s Antigravity docs. Last checked: October 29, 2026.

On this page

What Google Antigravity is

Google calls Antigravity its “agentic development platform” (antigravity.google). The overview says it evolved from the Agent Manager into a standalone desktop command center and now ships these surfaces:

Google says the App and the CLI “run the same agent, so most features work on both” (feature overview). Models named on the overview include Gemini 4 Argon, Gemini 3.8 Flash and Gemini 3.1 Pro, with a separate limit meter for Claude and GPT models (overview).

How you install it

From Google’s Get started guide:

• App: download the installer for macOS 12 or later, Windows 10 (64-bit) or 11, or Linux (glibc 2.28 or later); it comes pre-installed on Googlebooks.
• CLI on macOS, Linux or Googlebook:

curl -fsSL https://antigravity.google/cli/install.sh | bash

On Windows PowerShell:

irm https://antigravity.google/cli/install.ps1 | iex

Then check the install and start a session in a project folder:

agy --version
agy

• Editors: install the Google Antigravity extension from the VS Code Marketplace (VS Code 1.90 or later), JetBrains AI > Settings > Agents, the Visual Studio Marketplace, Zed’s External Agents registry, Xcode’s Intelligence settings or the Eclipse Marketplace (public preview).
• SDK in a Python 3.10+ virtual environment:

pip install google-antigravity

If you are coming from Gemini CLI, see migrating from Gemini CLI to Antigravity CLI.

What the agent can do

Google’s overview and feature overview list these capabilities:

• Projects: agents work in projects that define which folders they can access and which settings and permissions apply. The App supports Git worktrees, so an agent can run in an isolated worktree instead of your active folders.
• Subagents: the main agent can hand research, builds or test runs to subagents, each with a clean context, running in parallel with your conversation.
• Browser subagent: the /browser command hands a web task to a subagent that drives Google Chrome through the Chrome DevTools MCP server and records each session as a WebM video (/browser docs). Google notes it isn’t available in the Antigravity IDE, which has its own browser features.
• Artifacts: implementation plans, multi-file diffs and browser recordings that you review and comment on; in the CLI, Ctrl + R opens the Artifact Review screen (Get started).
• Slash commands: /plan, /grill-me, /goal, /boost, /schedule and /teamwork-preview; Google says /boost and /teamwork-preview require a paid plan (feature overview).
• Remote Control: drive and monitor App sessions and CLI instances from a web browser on another device.

Skills, rules, MCP and hooks

According to the feature overview, the App and CLI share the same customization layer:

• Skills: folders of instructions, each with a SKILL.md file.
• Rules: persistent instructions and coding standards for every conversation.
• MCP servers: local tools, databases and remote APIs through the Model Context Protocol.
• JSON hooks: local shell scripts that run before tool calls, after model responses or at stopping conditions.
• Plugins: namespaced bundles of skills, agents, rules, MCP servers and hooks; plugins installed in the App also appear in the CLI.

Permissions and safety defaults

Google says agents are “secure by default” (feature overview):

• In the App, agents ask before running terminal commands outside the isolated Terminal sandbox and can only read and write inside the project’s folders. Google warns: “If you broaden your permission settings (for example, by choosing the Turbo preset), the agent has read and write access over your full machine.” Settings are scoped per project, and the docs give the example of a more permissive setting for a trusted project and a more restrictive one for an untrusted folder.
• In the CLI, the default request-review permission asks before the agent writes files, runs shell commands or uses web tools unless an allow rule matches. The terminal sandbox is off by default; turn it on with enableTerminalSandbox in settings.json or the --sandbox flag. The agent also asks before reading or editing files outside your workspace folders.

Plans and pricing

Google’s pricing page lists a $0/month individual plan with access to a variety of agent models, unlimited Tab completions, unlimited Command requests and “basic weekly rate limits”, and paid tiers with more rate limits and a flexible AI credit pool. The overview names the paid individual plans as Google AI Plus, Pro and Ultra, and the pricing page offers a business plan via Google Cloud; the overview lists billing through Gemini Enterprise licenses or pay-as-you-go through the Gemini Enterprise Agent Platform API. What each tier includes is tracked in the plans tracker, and the free tier’s conditions are covered in is Google Antigravity free.

For organizations, the overview lists Google Cloud SSO and Workforce Identity Federation, US and EU data residency, VPC Service Controls and request logging, and says enterprise use is governed by the Google Cloud Terms of Service.

How it compares

For side-by-side comparisons built from both vendors’ docs, see Antigravity vs Claude Code and Google Antigravity vs Cursor.