Antigravity CLI Prompting Guide: Google's Gemini CLI Replacement (2026)

If you went looking for Gemini CLI recently and found conflicting answers, here is the resolution: Google retired Gemini CLI for consumer accounts on June 18, 2026, one month after announcing the transition at I/O on May 19. Google AI Pro and Ultra subscribers, the free individual tier of Gemini Code Assist, and GitHub organizations on standard Gemini Code Assist all lost access on that date. Organizations holding Standard or Enterprise licenses kept it, using paid API keys.

The replacement is Antigravity CLI, invoked as agy. It is a Go rewrite rather than a patch, it shares an agent harness with Google's Antigravity desktop application, and it is built around parallel agents rather than a single conversational loop.

The open-source gemini-cli repository is still publishing builds, which is why a quick GitHub check makes the project look healthy. Repository activity is not entitlement: if you are on a consumer plan, the CLI no longer serves your requests regardless of what the release page says.

Key takeaways:

• Gemini CLI is gone for consumer tiers. Enterprise licenses are the exception, not the rule.
• Migration is cheap: GEMINI.md still works unchanged, and agy also reads AGENTS.md.
• Extensions come across as plugins; skills and hooks carried over.
• Three execution modes replace the old approval prompt-per-action rhythm: default, accept-edits, and plan.
• Parallel subagents are the genuinely new capability, and they reward disjoint scoping and punish vague prompts.
• Skills live at .agents/skills/<name>.md and become /name slash commands.
• MCP configuration moved to a workspace file, and remote server entries need a field rename when ported.

Installing and Migrating

Install on macOS or Linux:

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

On Windows, PowerShell:

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

agy authenticates through your system keyring, falling back to Google Sign-In. For CI, recent versions support a keyless path by setting the model provider in settings and supplying an API key through the environment, which avoids interactive login in a pipeline.

Bring your old configuration across:

agy plugin import gemini    # Gemini CLI extensions become Antigravity plugins

Then check what actually loaded with /config. Settings live under your Gemini config directory, in an antigravity-cli/settings.json file, so a stale value in the old location is a common cause of "my setting is being ignored".

Models

Antigravity CLI runs Google's current lineup, with Gemini 3.1 Pro as the flagship and Flash-tier models for faster, cheaper work, and it can also drive selected third-party models. Switch with --model at launch or /model mid-session.

The pattern that pays off is the same one that pays off everywhere: match the model to the task rather than pinning the strongest one for the whole day. Planning a migration, reasoning about an unfamiliar subsystem, and debugging something subtle justify the flagship. Applying a decision you have already made does not. For the model-by-model view, see which AI model for coding.

The Three Modes

Modes are set with --mode at launch, cycled with Shift+Tab during a session, or persisted in settings:

There is also a flag that skips permission prompts entirely. It exists for unattended automation and it deserves the same treatment as its equivalents in other agents: use it inside a disposable worktree, on a branch you can throw away, and never on a task where the agent will read untrusted text. See prompt injection defense for why that last condition matters more than it sounds.

Prompting for Parallel Subagents

This is where Antigravity CLI diverges from the single-loop agents. Subagents are dispatched to work in parallel, report progress in the status bar, and post their diffs back into the conversation when they finish. /agents opens a panel to monitor them.

Parallelism changes what a good prompt looks like. A single-loop agent given a fuzzy scope wastes time. Three parallel agents given a fuzzy scope produce conflicting edits to the same files, and you pay for all three.

The rules that make parallel delegation work:

• Partition by path, not by topic. "Error handling" and "validation" overlap in the same route files; src/api/errors/** and src/api/validation/** do not.
• State the exclusions. "Do not touch route handlers" prevents the most expensive class of collision.
• Give each workstream its own definition of done. If each one must independently leave the test suite green, you can merge them in any order.
• Keep the count honest. Two well-scoped subagents beat five overlapping ones, and reviewing five parallel diffs is its own cost.

This is multi-agent orchestration with the coordination pushed onto you, which is a reasonable trade as long as you actually do the coordinating. For the broader patterns, see the multi-agent prompting guide.

Context Files: GEMINI.md and AGENTS.md

agy reads GEMINI.md unchanged, and it also reads AGENTS.md. That leaves you a decision rather than a migration:

• Only Google tooling touches this repo? Keep GEMINI.md.
• More than one agent works here? Consolidate into AGENTS.md so Codex CLI, Claude Code, Cursor, and everything else read the same rules, and keep GEMINI.md for genuinely Google-specific instructions only.

The second case is the common one in 2026, and split-brain instruction files are a real failure mode: two files, two sets of conventions, and an agent that follows whichever one it happened to read. The AGENTS.md guide covers the layering rules.

Skills

A skill is a markdown file under .agents/skills/. The file name becomes the command:

.agents/skills/
  ship-check.md      →  /ship-check
  db-migrate.md      →  /db-migrate

Skills can declare which tools they are allowed to use and can define nested subagents, which makes them the natural place to encode a procedure that fans out. Files that lived under .gemini/skills/ are imported during migration.

The design rules transfer from Agent Skills generally: the description determines whether the skill is ever selected, the body should read as a procedure rather than a description of one, and a skill that duplicates what is already in your context file is a skill that will never earn its tokens. The Agent Skills guide covers this in depth.

Useful Session Commands

/grill-me deserves a note. It inverts the usual failure mode, where an agent guesses at an ambiguous requirement and you discover the guess in the diff. Forcing clarification up front costs one round trip and saves a run. If you find yourself reaching for it constantly, that is a signal your AGENTS.md is missing the context you keep having to supply by hand.

/browser deserves a warning. Once an agent can browse, page content becomes input to a system that also has file and shell access, which is the exact configuration that makes indirect prompt injection dangerous. See the browser agents guide for the threat model.

What This Means Beyond Google

The retirement of Gemini CLI a year after launch is worth reading as a signal rather than a one-off. The terminal agents are converging on the same shape: OS-level sandboxing, a graduated permission model, background subagents, markdown-defined skills, MCP for tool reach, and AGENTS.md as the shared instruction file. Codex CLI and Claude Code have the same feature list under different names.

The practical implication for your prompting is reassuring: the transferable skill is not tool-specific syntax. It is scoping a task tightly, naming the verification command, writing durable rules into a file instead of retyping them, and choosing a permission level that matches what you are willing to have go wrong. Those survive the next rename.

Where to Go Next

• The category pillar: prompting AI coding agents
• The cross-tool instruction file: AGENTS.md guide
• The other terminal agent worth knowing: Codex CLI prompting guide
• Prompting Google's models generally: Gemini prompting guide
• Build reusable task prompts with the SurePrompts generator