Skip to content
A

Agent harnesses

Antigravity CLI

Google's terminal agent harness with subagents and skills

Proprietary

Antigravity CLI (agy) is Google's terminal client for its Antigravity agents. It runs parallel subagents, slash commands, MCP servers, plugins, skills and hooks from the shell, and drives Gemini models through Google OAuth or a Google Cloud project. Closed source, with print mode, structured output and an ACP server for wiring it into other harnesses.

What Antigravity CLI is

Antigravity CLI is Google’s terminal surface for its Antigravity agents. The binary is agy, and everything an agent needs runs without leaving the shell: natural-language tasking, parallel subagents, slash commands, MCP servers, plugins, skills and hooks.

Key features

  • Subagents - delegate background tasks to concurrent sessions. Type /agents for the panel, ctrl+k approves tools without leaving the keyboard.
  • Slash commands - reach plugins, MCP servers, skills and hook configuration from the prompt.
  • Small footprint - the lightest way to invoke and monitor Antigravity agents.
  • Print mode - -p runs a single prompt non-interactively, with --output-format stream-json and --json-schema for structured output.
  • Effort control - --effort low|medium|high|xhigh|max sets reasoning depth per session.
  • Modes - --mode plan|auto|accept-edits, and --sandbox for terminal restrictions.
  • ACP server - the acp subcommand speaks the Agent Client Protocol over stdio, so other harnesses can drive it.

Install

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

The default binary lands at ~/.local/bin/agy. macOS, Linux and Googlebook are supported through the shell installer; Windows uses PowerShell or CMD. arm64 Linux is a first-class target.

Sign-in

The CLI refuses to list models until you authenticate, and there are two paths: Google OAuth, or a Google Cloud project. The OAuth request asks for https://www.googleapis.com/auth/cloud-platform among its scopes, which is broad access to your Cloud projects rather than a narrow model scope. A dedicated account is the sensible choice.

What it does not do

Antigravity CLI is closed source, and it drives Gemini models rather than a local checkpoint. Nothing here runs a model on your own GPU; the local software is the harness, not the intelligence.

Why the harness matters

Google’s own Gemini 4 Neon post reports the effect of this harness directly: on SWE-sweep the same model scores 12.8% in a bare mini-SWE harness and 39.0% inside Antigravity. Nearly three times the result, with no change to the weights. That number is the argument for caring about harnesses at all.

At a glance

Who's behind it

Google

Source
closed-source
License
Proprietary
Platforms
cli
Install
curl -fsSL https://antigravity.google/cli/install.sh | bash
Model support
Gemini models, including Gemini 4 Neon, through Google OAuth or a Google Cloud project. No local model support.