Antigravity CLI

Starter Google's terminal coding agent and the successor to Gemini CLI. Binary: agy.

Checked against agy 1.5.0. Flags change quickly, so confirm with agy --help.

Source of truth

Docs: antigravity.google/docs
Changelog: GitHub, or run agy changelog

Install

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

agy --version
agy

The first launch opens a browser sign-in, then keeps your credentials in the OS keyring. On a remote machine over SSH, it prints a URL and you paste the code back. For an API key, see the authentication docs.

Moving from Gemini CLI

Google stopped serving Gemini CLI on Google AI Pro, Ultra and free individual plans from June 18, 2026. Paid API keys and enterprise access were not affected. Google says there's no 1:1 feature parity yet.

The first launch offers a checklist of what to migrate. Then:

agy plugin import gemini

What	Where it goes now
Context files	No change: GEMINI.md, AGENTS.md and ~/.gemini/GEMINI.md still work
Extensions	Plugins, through the import above
Global skills	Stay in ~/.gemini/skills/
Workspace skills	Move by hand from .gemini/skills/ to .agents/skills/
Global MCP servers	~/.gemini/config/mcp_config.json
Workspace MCP servers	.agents/mcp_config.json
Remote MCP config	Use serverUrl instead of url or httpUrl

Coming from Claude Code? agy plugin import claude brings your plugins across.

Everyday flags

Flag	What it does
-c, --continue	Continue the latest conversation
--conversation ID	Resume a specific conversation
-p "..."	Run one prompt and print the result
-i "..."	Run a prompt, then stay interactive
--mode plan	Plan mode. Also accept-edits and auto
--model, --effort	Choose the model and reasoning effort
--agent NAME	Use a specific agent
--add-dir PATH	Add a folder to the workspace
--sandbox	Run with terminal restrictions on
--output-format json	Machine-readable output for scripts

Auto-approving

--dangerously-skip-permissions approves every tool call with no prompt. Prefer --sandbox, or --mode auto, which needs the sandbox and isn't available on business or enterprise accounts.

Slash commands worth learning

Command	What it does
/goal	Keep working toward a goal
/plan	Plan before changing anything
/schedule <when> <task>	One-off timer or cron schedule. Scheduled agents use a fixed model
/agents, /tasks	Switch agents and see tasks
/browser <task>	Hand a task to a browser subagent
/btw	Ask a side question without derailing the turn
/diff, /rewind	Review changes, or go back
/fork, /resume	Branch or reopen a conversation
/skills, /plugin, /mcp, /hooks	Manage extensions
/permissions	Review what's allowed
/voice	/voice live to talk, /voice dictate to dictate. F5 also dictates

Headless and scripts

agy -p "summarize this repo" --output-format json

The JSON result includes finished_reason when a turn ends early (print_timeout, interrupted or error), so scripts can spot a partial answer. When --print-timeout cuts a run short, it exits with code 3.

Hooks

PreInvocation hooks run before the model is called, and returning "terminate" from terminationBehavior stops the turn, for example to block a prompt containing sensitive data. /hooks lists what's active.

Secrets

• agy doesn't load GEMINI_API_KEY from .env files, so export it in your shell or use the keyring sign-in.
• Settings include permissions allow, ask and deny rules, but I found no .env-specific guidance in the CLI reference. Test any rule on your own setup.
• The dependable fix is to keep real secrets out of the project. See Keep .env away from agents.

Related

• Agent CLI game changers
• Gemini CLI (legacy)
• AI agent CLIs