Google Antigravity is Google’s AI coding agent, sold three ways: the Antigravity IDE, an editor with an agent manager that launched in November 2025; the Antigravity CLI, a terminal agent called agy that replaced Gemini CLI for personal Google accounts on 18 June 2026; and Antigravity 2.0, a desktop app from I/O 2026 that runs several agents on one codebase in parallel. All three run Gemini plus a fixed list of other models that includes Claude, sign in with a Google AI plan (Free, Pro or Ultra) or a Gemini API key, and share the same permission modes and sandbox. If you had Gemini CLI on a personal account, Antigravity CLI is where Google sent you.
The three Antigravity products
| Antigravity IDE | Antigravity CLI (agy) | Antigravity 2.0 | |
|---|---|---|---|
| What it is | An editor with a built-in agent and an agent manager | A terminal agent, written in Go, that reasons over the project, edits with permission and runs commands | A standalone desktop app that orchestrates several agents on parallel tasks |
| Since | November 2025 | 2026; Gemini CLI’s successor for personal accounts from 18 June 2026 | Announced 19 May 2026 at I/O |
| Where | macOS, Windows, Linux | Terminal; editors through ACP | Desktop |
| Enterprise | Not supported for enterprise accounts as of October 2026 | Yes | Yes |
Google is moving the multi-agent work out of the IDE: an upcoming IDE release removes its agent manager, and 2.0 takes that job. If you’re starting now, the CLI and 2.0 are the products with a future; the IDE is the one being narrowed. There is also an Antigravity SDK, the same harness as a library, for building agents that run on your own infrastructure.
Models, plans and sign-in
- Models: Gemini, plus a fixed list Google maintains that includes Claude models. You pick from the list; you can’t add a provider or a local model to the CLI. The SDK is the exception.
- Sign-in: a Google account on the Free, Pro or Ultra plan, or a Gemini API key. Pro and Ultra get a quota that refreshes every five hours; the plan comparison, next to ChatGPT, Claude and Copilot, is in AI coding subscriptions compared.
- Gemini CLI: still exists, for Gemini Code Assist Standard and Enterprise and paid API keys only. Google says Antigravity CLI does not match it feature for feature yet; if you scripted Gemini CLI in CI, read Google’s migration notes before switching.
Permissions and sandbox
Antigravity CLI has four modes, default, accept-edits, plan and auto, from asking before every edit and command to running without asking at all. Under the modes sit allow, ask and deny rules per tool, the same shape as Claude Code’s. The sandbox runs on Linux, macOS and natively on Windows 11 24H2 with a 2026 update, which puts Antigravity in the short list of agents with a native Windows sandbox, next to Codex. How to sandbox AI coding agents shows what each one restricts.
auto is the YOLO mode. Google’s docs, like everyone’s, say to use it in a container. What each agent’s skip-all flag does, and why people reach for it, is in YOLO mode.
Install the CLI
Windows, in PowerShell:
irm https://antigravity.google/cli/install.ps1 | iexmacOS and Linux:
curl -fsSL https://antigravity.google/cli/install.sh | bashThen run agy in your project and sign in with Google or paste an API key. The setup next to Claude Code, Codex and OpenCode, including the permission settings for each, is in terminal coding agents.
What Antigravity doesn’t do
- No open-source license. The CLI, IDE and 2.0 are under Google’s terms. Gemini CLI is Apache-2.0, but it’s now enterprise-only.
- No model choice beyond the list. OpenCode, Cline and Kilo Code run any provider; Antigravity runs Google’s list.
- No local models in the CLI. A Gemma model through Ollama works with OpenCode or Codex, not with
agy. - No benchmark row. Antigravity has no entry on Terminal-Bench 4.0, so it can’t be ranked against Claude Code and Codex on the same tasks.
Antigravity alternatives
| Tool | Models | Pick it if |
|---|---|---|
| Claude Code | Claude only | You want the top benchmark rows and have a Claude plan |
| Codex | OpenAI; local with --oss | You pay for ChatGPT, or you’re on Windows and want a sandbox; open source |
| OpenCode | 75+ providers, local models | You want Gemini through an API key and any other model too, open source |
| Cursor | Cursor-hosted models | You want the editor more than the agent |
| Contenox | ChatGPT plan, API keys including Gemini, Ollama, vLLM; hosts Claude Code | You want permission rules you can read, per project, in a desktop app on Windows |
The full matrix, 16 agents on license, models, sign-in, Windows, permissions and sandbox, is in the AI coding harness comparison.
Contenox
Contenox takes a Gemini API key like any other provider and runs the agent under rules you can read in Permissions, instead of a mode switch: builds and tests allowed, edits and risky commands asking, secrets blocked. It hosts Claude Code when you have a Claude plan, under the same rules. Download it here.
Related guides
- Set up a Gemini coding agent: AI Studio keys, free-tier data use, 429 fixes
- Terminal coding agents: Antigravity CLI next to Claude Code, Codex and OpenCode
- AI coding agents on Windows: which sandboxes exist there
- Best AI coding agents: where Antigravity sits in the ranking