Antigravity CLI (agy): Commands, Modes, and Auto-Approve

Antigravity CLI is Google's terminal coding agent, invoked with the agy command. It shipped as the replacement for Gemini CLI, which Google shut down for free, Pro, and Ultra users on June 18, 2026. If you typed gemini in your terminal, agy is what Google wants you typing now.

This is the practical guide to the tool itself: what it is, how to install it, the commands that matter, and the honest comparison against Claude Code. If your goal is specifically moving off the dead Gemini CLI, read the Gemini CLI shutdown and migration guide first, then come back here for the day-to-day usage.

What Is Antigravity CLI?

Antigravity CLI is a Go-based binary that brings Google's agent-first workflow into the terminal. It's the command-line sibling of Google Antigravity, the agent IDE - same multi-agent model, no editor window.

Three things define it:

• Single Go binary. It ships as one compiled executable - install is a one-line script, no Node or Python runtime to manage. It picks a Flash-tier Gemini model by default, and you can override with /model in-session or the --model launch flag.
• Multi-agent by default. Like the Antigravity IDE, agy is built to dispatch and coordinate multiple agents rather than run one linear chat. It shares the same agent harness and settings as Antigravity 2.0.
• A Google-led product, not a community project. Gemini CLI was open source and took 6,000+ community pull requests before Google wound it down. agy is built and controlled by Google around that shared harness - don't expect the same community-extension flow.

One caveat worth knowing before you commit: at launch, Antigravity CLI does not have full 1:1 feature parity with Gemini CLI. Google kept the critical pieces (Agent Skills, Hooks, Subagents, and Extensions-now-plugins), but if you depended on a specific flag, check it works in agy before you rewrite your scripts.

How to Install Antigravity CLI

Install is a one-line script. On macOS or Linux:

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

On Windows (PowerShell):

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

Then start it and authenticate:

agy

agy authenticates through your system keyring, falling back to Google Sign-In (pick "Use a Google Cloud project" when prompted). Check your setup with /config, then smoke-test it non-interactively:

agy -p "summarize this repo"

Both agy and the old gemini command can coexist on the same machine, which makes a gradual cutover easier if you still have a paid API key keeping Gemini CLI alive.

The Slash Commands That Run It

agy is driven by slash commands in the TUI. These are the ones you'll actually use:

Two design changes worth flagging if you're coming from Gemini CLI. Browser use is now opt-in via /browser - Google moved it behind a command because the agent kept reaching for the browser when it shouldn't. And subagents are dynamic: you hand the orchestrator a goal in plain English and it decides what specialized subagents to spawn and runs them in parallel, no orchestration code from you.

/goal and /schedule are the pair worth sitting with. /goal keeps the current task running until it is completely finished, without stopping to ask you for input along the way. /schedule runs an instruction later as a one-time timer or on a recurring schedule - a separate scheduled task, not a repeat of the run you're in. That's the actual mechanics of loop engineering: deciding when an agent keeps going, when it stops, and when it starts again on its own, not just what to prompt it with.

On models, Antigravity's platform page lists Gemini 3.7, 3.6, and 3.5 Flash, Gemini 3.1 Pro, Claude Sonnet 4.6 and Claude Opus 4.6 (thinking variants), and GPT-OSS-120b as of late August 2026; what /model actually offers in agy depends on your account and plan. Skills work like Gemini CLI: drop a markdown file at .agents/skills/lint.md and it becomes /lint. Hooks are JSON lifecycle interceptors (before tool call, after file edit, on session start), with workspace hooks overriding global ones. Sandboxing uses nsjail on Linux and sandbox-exec on macOS.

Does agy Have a YOLO Mode?

No. There is no --yolo flag in agy. Gemini CLI had one, Antigravity CLI replaced it with two separate controls, and which one you reach for depends on whether you care about file writes or shell commands.

Execution modes handle file edits. There are three, and you set one at launch:

agy --mode=accept-edits

In accept-edits, the standard file operations (write_to_file, replace_file_content, multi_replace_file_content) run automatically. Press Shift+Tab inside the prompt box to cycle default, accept-edits, plan, and back without restarting the session. To make a mode your default, set agentMode in ~/.gemini/antigravity-cli/settings.json or open the settings editor with /config.

Shell commands are governed separately. Modes only cover edits. run_command still goes through your tool permission rules in every mode, and you manage those with /permissions. So switching to accept-edits will not on its own get you an unattended run: the first command the agent wants to execute still stops and waits.

The actual approve-everything switch is --dangerously-skip-permissions, which approves all tool calls including file writes and command execution:

agy -p "Run the test suite and report failures" --dangerously-skip-permissions

That is the closest thing to the old --yolo, and the flag name is the warning. Run it in a container or a sandbox, not on your laptop pointed at a repo you care about.

Migrating from Gemini CLI

If you're coming straight off Gemini CLI, the move is mostly mechanical and agy ships an importer:

• Import your extensions: agy plugin import gemini converts installed Gemini CLI extensions into Antigravity plugins.
• Move workspace skills: mv .gemini/skills/ .agents/skills/ - Antigravity looks for them under .agents/skills/.
• MCP config: Antigravity uses a dedicated mcp_config.json (workspace path: .agents/mcp_config.json). For remote servers, rename the url field to serverUrl. Release 1.1.16 (2026-08-20) added agy mcp add, remove, list, enable, and disable for the user-level MCP config; workspace-scoped servers still live in .agents/mcp_config.json and may require direct editing.
• Context files stay put: GEMINI.md keeps working - no rename needed. agy reads both GEMINI.md and AGENTS.md. New projects increasingly standardize on AGENTS.md, but that's optional.
• Update automation: swap every gemini call in CI/CD, shell scripts, and cron jobs for agy.

The one that bites people is automation. Any pipeline calling gemini -p "..." broke on June 18 with no grace period. Audit your scripts first - that's where the silent failures hide. Full checklist in the migration guide.

The Quota Trap (Read This Before You Commit)

This is the single biggest complaint from early agy users, and it's not a bug - it's how the tool works. Subagents burn quota in parallel. One plain-English goal can spin up a nested loop of 10+ model calls without asking, and the meter runs the whole time.

The numbers from the field are brutal. One paid Google AI Pro user reported being locked out with an "Individual quota reached" error after exactly two prompts in agy - the same Pro quota that survived 1,000+ code modifications in the old Gemini CLI. The quota is also shared across the Antigravity desktop app, CLI, and SDK, so running agents in parallel across projects drains it even faster.

Three habits keep you out of the lockout:

• Watch /usage constantly. Unlike Gemini CLI, the agent can't see its own remaining quota and won't throttle itself. You are the governor.
• Don't over-parallelize. Pro/Ultra quotas comfortably support 3 to 5 parallel subagents. Past that you hit the ceiling before lunch. Less parallelism often means more total throughput across the day.
• Reach for /grill-me on expensive work so the agent plans before it spends, instead of looping autonomously on a vague goal.

If you're on the free tier, expect aggressive throttling and queued subagents that never start. For real throughput, connect a billing-enabled Google Cloud project.

Scripting and CI Gotchas

If you automate agy, three traps will cost you a debugging session each:

• A bare agy call is not a script. Non-interactive runs need -p (aliases --print and --prompt), which runs the prompt once and exits. Approval is a separate flag: --dangerously-skip-permissions approves all tool calls, including file writes and command execution. There is no --headless flag and no --approve policy, so agy -p "..." --dangerously-skip-permissions inside a sandbox is the unattended shape.
• Auth has to exist before CI runs. By default headless mode uses cached credentials, so sign in once from an interactive agy session on that machine first. Since 1.1.13 (2026-08-14) there is a second path built for CI: set modelProvider: "gemini" in settings.json and export GEMINI_API_KEY, and agy runs against the Gemini API directly with no sign-in (a GOOGLE_GEMINI_BASE_URL override is available for custom endpoints). Note this restores the environment variable earlier releases ignored, so re-check any migration notes written before mid-August. An unauthenticated non-interactive run exits with an authentication required error rather than hanging.
• Pick an output format if a script reads the result. --output-format takes text (the default), json for a single object with metadata, or stream-json for newline-delimited events. Early versions could also silently drop stdout when -p ran under a pipe, subprocess, or redirect, so verify you actually get bytes back before trusting it.

None of these are dealbreakers, but they're the difference between a clean cutover and a broken Monday pipeline. Test on a throwaway workflow first.

Antigravity CLI vs Claude Code

This is the comparison most builders are actually searching for, because Claude Code is the closest terminal-native analog. Here's the honest split:

Short version: if you live in the Google ecosystem and want parallel agents from the terminal, agy is the natural pick. If you want the most mature terminal coding agent today and don't care which cloud it talks to, Claude Code is still the safer bet. Many builders will run both - agy for Gemini-backed tasks, Claude Code for everything else.

For the full field including Cursor and the Antigravity IDE, see our best AI coding agent comparison.

Should You Switch?

• You were on Gemini CLI free/Pro/Ultra: You have to move - the tool is dead. agy is the path of least resistance if your work is Google-centric.
• You hold an enterprise Gemini Code Assist license: You keep both tools, so there's no rush. Test agy on a side project before you standardize on it.
• You're tool-shopping fresh: Don't pick agy purely because it's new. The missing Gemini CLI parity and the quota trap are real trade-offs - if you run high-volume agent loops, the per-prompt cost ceiling will bite. Weigh it against Claude Code and Cursor on the work you actually do.

The smart play right now is to install agy, port one real workflow, and keep your fallback alive until the CLI proves it covers your cases. New tools earn trust by surviving your actual pipeline, not a toy demo.