Antigravity CLI: Google's terminal coding agent, explained

Published October 21, 2026 · by the AQ team

Antigravity CLI is Google's AI coding agent for the terminal: a single Go binary named agy that you run inside a project directory, where it reasons over the codebase, edits files with your permission, and runs commands, sharing its agent engine with the Antigravity 2.0 desktop app. It is the official successor to Gemini CLI: Google announced the transition at I/O on May 19, 2026, made the Antigravity CLI available to everyone the same day, and stopped serving Gemini CLI requests for Google AI Pro and Ultra subscribers and free individual users on June 18, 2026. This page covers the agy command, the migration, and the model lineup, as of October 21, 2026.

What happened to Gemini CLI

Per the Google Developers Blog post announcing the change ("An important update: Transitioning Gemini CLI to Antigravity CLI," May 19, 2026), Gemini CLI and the Gemini Code Assist IDE extensions stopped serving requests on June 18, 2026 for Google AI Pro and Ultra users and for free Gemini Code Assist for individuals users. Enterprise customers with a Gemini Code Assist Standard or Enterprise license kept their access, and Gemini CLI remains usable through paid Gemini API keys. For everyone else, the terminal path forward is the Antigravity CLI, part of Google consolidating its developer tools into one agent-first platform.

Google's own post says the new CLI keeps Gemini CLI's core features (Agent Skills, Hooks, Subagents, and Extensions, now called plugins) and adds asynchronous multi-agent background workflows, while being candid that there was not full feature parity at launch. Two structural changes matter beyond features. The runtime: Gemini CLI was a TypeScript npm package; the Antigravity CLI is a prebuilt Go binary with no Node.js requirement. And the license: Gemini CLI was open source under Apache 2.0, while the Antigravity CLI ships as a binary whose source Google has not published as of October 2026 (the google-antigravity GitHub repository hosts the README and issue tracker). That repository also states that using the CLI means agreeing to let Google collect Interactions data to improve the product, with an opt-out in settings: worth reading before pointing it at sensitive code.

Install and first run

Installation is a one-line script per platform, verified on Google's repository and codelab as of October 2026:

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

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

# Then, in a project
cd your-project
agy

On first launch, agy asks you to sign in with Google OAuth or a Google Cloud project, opens your browser, and has you paste a code back into the terminal, then walks through a color theme, the terms, and folder trust. Over SSH it prints an authorization URL you complete on another machine, which matters because remote boxes are where a terminal agent earns its keep. Credentials live in the system keyring and /logout clears them. Settings are a JSON file under the home directory (still under ~/.gemini, a deliberate continuity choice); Google's codelab recommends editing them through /config rather than by hand, since a malformed file can stop startup.

What carries over from Gemini CLI

Google's migration guide is specific about what moves and what needs a hand, as of October 2026:

• Context files need no changes. A GEMINI.md or AGENTS.md in the workspace keeps working as is, and so does a global ~/.gemini/GEMINI.md.
• Extensions become plugins. One command, agy plugin import gemini, converts Gemini CLI extensions into native plugins: skills migrate, legacy custom commands are converted to skills, and MCP server definitions come along.
• Workspace skills move. Skills in .gemini/skills/ must be moved to .agents/skills/ before the agent picks them up as slash commands.
• MCP configuration moves out of settings. Server definitions leave the old inline settings.json for a dedicated mcp_config.json (global under ~/.gemini/config/, per-workspace under .agents/), with the url or httpUrl key renamed to serverUrl.
• Auth migrates. Active session tokens move into the OS keyring on first launch.

Models, plans, and limits

Model access is tiered by Google AI plan. The table below is Google's own availability matrix as of October 21, 2026; Google has announced that the older third-party rows come out of the lineup on November 2, 2026, so treat this as a snapshot.

Yes, Claude in a Google CLI: Google serves third-party models inside Antigravity on the paid plans, with usage tracked separately (one five-hour and weekly quota for Gemini models, another for Claude and GPT models; /usage shows both). Models run at Low, Medium, or High thinking levels, some Flash models offer a Fast tier, /effort changes the level, and /boost invokes a multi-agent deep reasoning mode. Google's plans documentation describes quotas qualitatively rather than in numbers: non-subscribers get a quota that refreshes weekly, Pro refreshes a larger quota every five hours until a weekly cap, Ultra gets the most generous limits, and Pro and Ultra can buy AI credits for overages (/credits tracks them). A model choice persists for the conversation, and switching mid-run takes effect after the current turn finishes. There is no bring-your-own-key option.

What it is good at

The architectural bet is asynchronous multi-agent work. The main agent can spawn subagents that run in parallel in the background (research, builds, test runs), surfaced in an /agents panel that shows each one's status and opens its full conversation, with keyboard shortcuts to approve a pending permission without leaving the main thread. On paid plans, /teamwork-preview assembles a multi-agent team for long-horizon projects. Around that core sit the extensibility surfaces: plugins bundle skills, rules, MCP servers, and hooks, installing from a marketplace, a GitHub URL, or a local path. Guardrails are configurable: the default request-review mode asks before commands and file writes, proceed-in-sandbox auto-runs commands inside OS-native isolation (nsjail on Linux, sandbox-exec on macOS, AppContainer on Windows), always-proceed is fully autonomous, and strict allows only read tools. Sessions keep persistent history with /resume to switch between them and /rewind to roll back to a checkpoint, agy -p runs a single prompt headlessly for scripts, and a session can be exported to the Antigravity 2.0 desktop app, which shares the same harness and syncs settings both ways.

Among coding harnesses it is the clear pick if your models come from a Google AI plan, and the bundled Claude access makes it more model-plural than its predecessor ever was. The tradeoffs: the closed source binary, the Interactions data collection default, and parity gaps Google itself acknowledged. The coding agent harness directory keeps the whole field in one place.

The machine problem

Like every terminal harness, the agy process runs where you started it. Persistent history and /resume make sessions recoverable on that machine, but close the laptop and the agent, its subagents, and its background work all stop until you come back; a session on your laptop is not resumable from your desktop. The vendor-neutral fix is the same as for Claude Code or Codex: run the CLI on a machine that does not sleep, inside tmux, and reconnect from wherever you are. Keep Claude Code running after closing your laptop walks through the patterns, and every one applies to agy unchanged.

The second gap: a session is single-player. Subagents parallelize the work, but no teammate can watch the session or steer it. That solo default is the industry pattern, not an Antigravity quirk: a July 2026 LeadDev analysis of 25,264 agent-generated pull requests across 2,361 popular GitHub repositories found that in 79 percent of agentic PRs the same developer both reviewed and modified the agent's contribution, and only about one in eight workflows involved multiple humans.

Where AQ fits

AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. The Antigravity CLI is one of the harnesses AQ runs: agy runs as the real CLI in a persistent tmux session on your team's VM, signed in with your own Google account (AQ never proxies or marks up usage on your own subscriptions), streamed live to the browser. The session survives closed laptops and resumes from any device, and a teammate can open the same workspace and watch or drive the same live session (typing into someone else's terminal is delegated: the owner approves a control request in one click).

Each workspace is an isolated git worktree on a branch like ai/1234-fix-auth, so agy in one workspace next to Claude Code in another never collides; agents commit, push, and open PRs with per-user GitHub auth, PRs are tracked per workspace, and labeling a Linear issue ai-task creates a workspace with ownership following the assignee. Pricing is two plans: Free is a personal sandbox for one person (AQ creates a private machine in an isolated network, nothing to install, no time limit), and the Team plan is $50 per user per month early access (standard $200, billed monthly), covering VMs you connect from your own cloud or a dedicated always-on AQ-managed VM, rate locked for your first 12 months.

Plainly: agy gives one developer a capable multi-agent terminal on one machine. If you want that terminal to outlive your laptop and be visible to your team, put it on a persistent shared machine.