Google Antigravity

Antigravity is Google's agent-first development platform. It ships in two forms that share a single agent engine: a desktop application and agy, a terminal client. Since June 2026 it is also the only supported path for individual developers using Google's coding agents — it absorbed both Gemini CLI and the Gemini Code Assist IDE extensions.

Overview

Antigravity launched on 18 November 2025 alongside Gemini 3. It was pitched as a different level of abstraction: rather than completing lines, you describe a goal and agents plan, edit across files, drive a browser, and report back with artifacts.

The strategically important thing happened later. At I/O on 19 May 2026 Google announced it was consolidating developer tooling under the Antigravity brand and retiring the standalone Gemini CLI and Gemini Code Assist IDE extensions. On 18 June 2026 Gemini CLI stopped serving requests from personal, free, and Google AI Pro/Ultra accounts. Enterprises with Gemini Code Assist Standard or Enterprise licences kept their access; every other developer was moved.

That makes Antigravity the answer to a question many people are still asking as "how do I use Gemini CLI" — the tool they mean is now agy.

Note also that Antigravity is not Gemini-only. Claude Opus 4.6, Claude Sonnet 4.6 and GPT-OSS-120b are selectable on several plans, which is unusual for a first-party vendor tool and makes Antigravity a genuine multi-model client rather than a distribution channel for one family.

Key Features

• Two surfaces, one agent core. The desktop app and the agy TUI run the same reasoning engine, and conversations export between them. Start on a laptop, continue over SSH.
• Parallel subagents. Modular, specialised or blank-slate agents spawned programmatically by the main agent — not the single fixed browser subagent the first release shipped with.
• Hooks. JSON config that intercepts the agent lifecycle: before a tool call, after a model call, at loop-stop conditions. This is the extension point that makes team-wide policy enforceable rather than advisory.
• Skills as slash commands. Markdown files with name/description frontmatter in .agents/skills/, automatically surfaced as slash commands in the TUI.
• Plugins. Namespaced bundles packaging skills, subagents, lint rules, MCP definitions and hooks into one installable asset — plugin.json plus optional skills/, agents/, rules/ directories.
• Multi-model selection. Gemini tiers with Low/Medium/High reasoning options, plus Claude and open-weight models on eligible plans.
• Browser control and image generation. Nano Banana 2 handles UI mockups and diagrams inline; the browser subagent verifies what the code actually renders.
• Remote-friendly. Explicit support for SSH, tmux and remote terminals, including a device-code-style auth flow for headless machines.

How It Works

You give agy a goal rather than an instruction. It plans, then executes across files, running commands and consulting tools as it goes. Slash commands like /goal and /schedule make the planning and deferral explicit.

The part that distinguishes it from a single-threaded agent is subagent orchestration: the main agent can spawn specialised workers in parallel and merge their results. For a task that decomposes cleanly — audit twelve modules, port a pattern across a monorepo — this is the difference between one long serial run and a fan-out. This is multi-agent execution as a product feature rather than a framework you assemble yourself.

Around that loop sit the hooks. Because they fire at defined lifecycle points and are plain JSON, an organisation can require a lint step after every edit, or block a tool entirely, without asking developers to remember. Skills and plugins carry the same logic in the other direction: repeatable procedures, packaged and shared.

Context comes from GEMINI.md or AGENTS.md in the workspace — Antigravity reads both, which is why migration needs no edits to those files.

Technical Details

• CLI binary: agy, written in Go
• Install location: ~/.local/bin/ (Unix), C:\Users\<Username>\AppData\Local\agy\bin (Windows)
• Auth: system keyring first (Keychain, Secret Service, Credential Manager), browser sign-in otherwise; SSH sessions get a manual authorisation URL
• Context files: GEMINI.md and AGENTS.md, both honoured unchanged
• Global skills: ~/.gemini/antigravity-cli/skills/
• Workspace skills: .agents/skills/
• Plugin manifest: plugin.json, plus mcp_config.json and hooks.json
• Models on all plans: Gemini 3.6 Flash, Gemini 3.5 Flash (Low/Medium/High), Gemini 3.1 Pro (Low/High)
• Models on Free / AI Plus / AI Pro: Claude Sonnet 4.6 (thinking), Claude Opus 4.6 (thinking), GPT-OSS-120b

Use Cases

• Migrating off Gemini CLI. For most individual developers this is the immediate reason to install it — the old command no longer works.
• Parallel repository work. Audits, framework upgrades and cross-cutting refactors that decompose into independent units are what subagents are for.
• Team policy in code. Hooks let a platform team enforce a review or lint step at the agent level, applied identically to everyone.
• Model comparison inside one workflow. Running Gemini 3.1 Pro and Claude Opus 4.6 against the same task without changing tools is a real advantage over single-vendor clients.
• UI work with visual verification. Nano Banana 2 for mockups plus browser control to check the rendered result closes a loop that text-only agents cannot.
• Remote and headless development. SSH-aware auth and a TUI make it usable on a server, where the desktop app is not.

Integrations

• MCP: server configuration via mcp_config.json, per plugin or globally
• Browser: built-in subagent for navigation and verification
• Editors: the desktop app is the editor; the CLI works alongside any of them
• Terminals: SSH, tmux, remote sessions
• Google AI plans: quota is drawn from your Google AI subscription rather than billed separately

Pricing & Access

Antigravity has no separate price. Access and quota ride on your Google AI subscription tier, and Google's own plans documentation deliberately does not publish exact request numbers — it describes quota shape rather than quantity.

Two caveats worth reading before committing:

The free tier has been cut repeatedly. It launched generous, dropped to roughly 20 requests per day by December 2025, and Google's docs state plainly that baseline limits exist "to the degree we have capacity". Google has not committed to a permanent free tier. Do not build a workflow that assumes today's free allowance.

Credits exist for overflow. Introduced in March 2026 at $0.01 each, sold in bulk (around $199 for 20,000). Google has not published what one credit buys in tokens or requests, which makes cost forecasting difficult.

Limits have moved in both directions: in May 2026 Google permanently tripled request limits on paid tiers and re-credited weekly quotas after developer complaints. The direction of travel is clearly toward metered usage.

Getting Started

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

# Windows (PowerShell)
irm https://antigravity.google/cli/install.ps1 | iex

Then run agy. On first launch it checks the keyring, opens a browser to sign in if needed, and — if it finds a Gemini CLI configuration — offers an interactive checklist of what to migrate.

Migrating from Gemini CLI:

agy plugin import gemini     # convert legacy extensions to plugins
agy plugin list              # confirm what came across

What carries over automatically: GEMINI.md and AGENTS.md unchanged, extensions as plugins, session tokens into the OS keyring, default visual settings.

What does not: custom terminal themes and experimental visual overlays, and — the one that silently breaks things — workspace skills. Move .gemini/skills/ to .agents/skills/ yourself, and global skills from ~/.gemini/skills/ to ~/.gemini/antigravity-cli/skills/. Nothing warns you; the skills simply do not appear.

Limitations

• Quota is the dominant complaint. Agent tasks stop mid-run when the window empties, and on the free tier the refresh is weekly. This is the most common reason people abandon the product, and no amount of model quality compensates for an agent that halts halfway through a refactor.
• Costs are hard to forecast. Google publishes neither exact rate limits nor the token value of a credit. You learn your throughput empirically.
• A forced migration is still a migration. The skills-directory move is manual and silent, and teams with heavy Gemini CLI customisation should budget real time rather than the fifteen minutes the migration guides promise.
• The free tier's future is explicitly uncommitted. Google's own wording ties it to spare capacity.
• Younger than its competitors' maturity suggests. The platform is under nine months old and absorbed two products in that time; docs and behaviour still move between releases.

Alternatives

• Claude Code — the most direct competitor for terminal-first agentic work, with a predictable subscription and no forced-migration history. Anthropic models only.
• OpenAI Codex — Apache-2.0 client, OS-level sandbox, GPT-5.6 models, and a published rate card. Better if you want to read the client's source or pin costs to tokens.
• Cursor — the mature agentic IDE, multi-model, and the closest analogue to the Antigravity desktop app.
• GitHub Copilot — the default for GitHub-centric teams, now agentic, with enterprise controls that predate everyone else's.
• Google Gemini — if you want the models without the agent harness.

Community & Support

• Documentation: antigravity.google/docs — the CLI, plugins and migration guides are the useful sections
• Blog: antigravity.google/blog — where plan changes and feature deep-dives are announced, including the ones that change your quota
• Migration guide: antigravity.google/docs/cli/gcli-migration

Because plan and quota changes have been frequent and consequential, the blog is worth following rather than checking after something breaks.