Antigravity CLI is the command-line member of the Google Antigravity
product family, alongside the Antigravity 2.0 desktop app, the Antigravity IDE and the
Antigravity SDK. It runs a coding agent inside a terminal user
interface and is invoked as agy. The product changelog lists its initial public release
as version 1.0.0; it dates 1.0.0 through 1.0.3 January 1, 2026, and the entries then run
from 1.0.4 (June 1, 2026) to 1.2.9 (September 23, 2026).
Google made the CLI available to everyone on May 19, 2026, in an announcement (An important update: Transitioning Gemini CLI to Antigravity CLI) that also made it the successor to Gemini CLI for consumer users. Google’s stated reason was that developers now want several agents communicating to split up complex work, which needs a terminal tool sharing one backend with the rest of the workflow. The announcement describes the CLI as built in Go for faster execution, as orchestrating multiple agents in the background so that large refactors or research tasks do not lock up the terminal session, and as sharing the same agent harness as the Antigravity 2.0 desktop application, so that improvements to the core agents apply wherever they are used. It carries over what Google called Gemini CLI’s most critical features — Agent Skills, hooks, subagents, and extensions, which become Antigravity plugins — while stating that feature parity would not be one-to-one at first.
The CLI shares its configuration with the rest of the family: from 1.0.5 its permissions
merge project-level rules, user settings shared with Antigravity, and the CLI’s own
settings.json, and customisations such as hooks, plugins and agents live in a shared
~/.gemini/config/ directory.
Capabilities
Execution modes. From 1.1.0 the default mode, request-review, pauses before each
file write to show a line-level diff in which the user can review, accept or reject individual code modifications, and
Shift+Tab cycles through default, accept-edits and plan. A /plan prefix replaced
the older /planning command in the same release. A proceed-in-sandbox permission mode
(1.0.1) auto-approves terminal commands that run inside the sandbox and asks only when a
command would leave it.
Headless use. Print mode (-p / --print) runs a single scripted turn. From 1.1.8 it
can emit text, json or stream-json (a typed NDJSON event stream), and --json-schema
enforces a schema on the structured result; from 1.1.15 --input-format stream-json keeps
one conversation open across newline-delimited prompts on stdin, so a driver program can
hold a session. The changelog frames these as serving CI, evaluation harnesses and scripts.
Several releases tighten headless behaviour: tools needing a permission the run lacks are
refused rather than hanging (1.1.3), refused actions are reported as denied_actions
(1.1.27), and from 1.2.6 a failing turn prints a structured AGY_ERROR line on stderr and
exits with code 3.
Delegation. The agent can hand work to subagents and background tasks, shown in a
status bar below the prompt (1.0.15) and managed from /agents and /tasks panels.
Custom agents can be defined as Markdown files with YAML frontmatter (1.1.6), selected at
launch with --agent (1.1.1), and given their own model tier (1.1.5). From 1.2.9,
an @ prompt syntax sends a message directly to a subagent conversation, with
autocomplete listing running and completed subagents.
The product documentation describes the CLI’s delegation model as asynchronous: rather
than locking the terminal during long builds, large codebase searches or multi-file
edits, the primary agent hands those operations to parallel subagents or background
tasks, so the user can keep prompting and inspecting files meanwhile. The /agents
panel lists every background agent with its role, state (running, done, killed or
error) and current step, and opens a detail view with that agent’s full reasoning log,
tool calls and outputs; /tasks tracks non-agentic background work such as shell
commands and test suites. Custom agents are discovered from .agents/agents/ in the
workspace and from ~/.gemini/config/agents/ globally. Because subagents may need
approval for a tool call, the CLI adds two shortcuts: Alt+J jumps straight to the next
subagent awaiting approval, and Ctrl+K approves the pending action shown above the
prompt without switching panels.
Extension points. The CLI loads skills, rules, plugins, lifecycle hooks and
Model Context Protocol servers; mcp subcommands (1.1.16) manage the
server list without hand-editing mcp_config.json.
Other features. Named in the changelog: a /btw command for side questions; /goal for long-running goals; a
/rewind command that reverts a conversation and its file edits to an earlier step;
/codesearch over the workspace; /voice dictation; an optional Vim editing mode
(1.1.11); and Remote Control, which lets another device follow and control a running
terminal session; releases from 1.1.19 onward mention it, and 1.2.6 added starting it per
session with --remote-control or /remote-control.
Adoption & Ecosystem
Sign-in options broadened over the release series: personal (consumer) accounts, a GEMINI_API_KEY
that runs the CLI directly against the Gemini API (1.1.13), Business sign-in for Gemini
Enterprise accounts, Workforce Identity Federation and Application Default Credentials
(1.1.10). A remote-control subcommand (1.2.0) registers the CLI with the operating
system’s service manager as a background daemon that persists across logouts and reboots.