Antigravity

Google Antigravity is an agent platform with four surfaces: an IDE, a terminal CLI, a desktop app for running agents in parallel, and an SDK. Individuals get a $0 plan with weekly rate limits; Google AI Pro and Ultra add higher limits and newer models, including Claude. Checked 31 October 2026.

At a glance

What it is

Google's own description of the products on its homepage is short:

• Antigravity App: "Your command center to manage multiple local agents in parallel."
• Antigravity CLI: "The lightweight, fast, terminal-first surface to work with Antigravity agents."
• Antigravity IDE: "The fully-featured, agentic IDE."
• Antigravity SDK: "Prototype custom agents leveraging Antigravity's harness with minimal code."

There is also an Extensions entry in the navigation, with no description we could read. The shared idea is a single agent harness reached from different surfaces, which is the same pattern Claude Code and Codex use. The difference is that Google puts the manager-of-many-agents view in its own app.

Antigravity also took over from Gemini CLI. On 18 June 2026, free individual access and Google AI Pro and Ultra access moved to Antigravity CLI. If you used Gemini CLI as an individual, this is its successor; see the Gemini CLI page for the split.

Install

The CLI download page (version 1.4.10 when we checked) gives these commands:

# macOS, Linux, Googlebook (Linux terminal)
curl -fsSL https://antigravity.google/cli/install.sh | bash

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

# Windows CMD
curl -fsSL https://antigravity.google/cli/install.cmd -o install.cmd && install.cmd && del install.cmd

The migration page names the command agy ("when you execute agy for the first time"), and the Gemini CLI README uses the same name, so that is what to type after installing.

The app and IDE are normal desktop downloads, with separate builds for Apple Silicon and Intel on macOS. The "minimum requirements" on the download page, such as macOS 12 (Monterey) or later, apply to the app, not the CLI.

First launch of the CLI

Open a fresh terminal, go to your project directory and run the launcher. On first launch a terminal setup asks about your color scheme, rendering mode and workspace trust. Useful commands from the CLI page:

Subagents are described as a way to "delegate background tasks to concurrent agent sessions." Slash commands also reach plugins, MCP, skills and hooks configuration, though the page gives no detail beyond that.

Plans and prices

The pricing page, checked 31 October 2026:

Rate limits are not given as numbers on that page; each links to an FAQ.

Models

The models page lists what the selector offers, with thinking levels where they exist.

Claude and GPT-OSS are labeled partner models. The selector shows separate weekly and five-hour limits for Gemini models and for Claude and GPT models. The point to take away: a free account gets Gemini 3.8 Flash and, until 2 November, two older partner options. The newest Claude models need a paid Google plan.

That second fact is the most unusual one on this page. Antigravity is one of the few ways to use Anthropic models without an Anthropic plan. It is not the same as Claude Code, which brings its own tools, but if the model is what you want, the plan matrix above is the thing to read.

Moving from Gemini CLI

Google publishes a migration page. The steps:

1. Run the Antigravity CLI for the first time. It detects legacy configuration and offers a checklist: extensions and global configuration, session tokens (moved to the OS keyring), and default settings.
2. Convert extensions with agy plugin import gemini. Legacy commands become skills, and MCP server definitions move to mcp_config.json.
3. Move workspace skills from .gemini/skills/ to .agents/skills/. The page says this is required.
4. Move MCP servers out of ~/.gemini/settings.json into a standalone mcp_config.json.
5. Change remote server keys url or httpUrl to serverUrl.

Custom terminal themes or experimental visual overlays from Gemini CLI may not carry over.

A first session, step by step

Everything here comes from Google's CLI and migration pages. We have not run it, so expect small differences in wording on screen.

1. Install and open a new terminal. Use the command for your platform from the install section, then start the CLI with agy inside a project folder.
2. Answer the setup prompts. The first launch asks about colour scheme, rendering mode and workspace trust. Say no to trust for a folder you did not write, since trust is what lets the agent act there.
3. Pick a model. Open the selector and look at what your plan lists. A free account should show Gemini 3.8 Flash, and Claude and GPT-OSS options that expire on 2 November 2026.
4. Check permissions. Type /config and read the permission settings before you give a real task. ctrl+k approves tools, so you will use it often at first.
5. Start small. Ask for a read-only task, such as explaining the folder layout, then a one-file edit with a test.
6. Try a subagent. The /agents panel is where background sessions show up. Give a long, independent job to one and watch how it reports back.

If your repository already has an AGENTS.md, the CLI reads it, so the project rules you wrote for another agent apply on day one.

Adding an MCP server

The migration page says global MCP configuration lives at ~/.gemini/config/mcp_config.json and workspace configuration at .agents/mcp_config.json. It also says remote servers use the key serverUrl, replacing url and httpUrl from Gemini CLI. From that, a remote HTTP server entry looks like this, with a placeholder address you replace:

{
  "mcpServers": {
    "docs": {
      "serverUrl": "https://example.com/mcp"
    }
  }
}

Two cautions follow. First, the format we can confirm covers remote HTTP servers only. If a server you need runs as a local process over stdio, check Google's current MCP page for how to declare it, because the pages we read do not show it. Second, keep servers per project where you can. A workspace file in .agents/ travels with the repository, so a teammate gets the same tools, while a global file follows only you. Wider guidance on choosing servers is on the MCP server list and the best MCP servers page.

The 2 November deadline, in practice

The model docs say Claude 4.6 models and GPT-OSS 120B are removed on 2 November 2026, which is two days after we checked these pages. Anyone on a free or Google AI Plus plan who relies on them has a short window.

Switching models mid-project can change style and tool use, so rerun your tests after you do. Pin the model name in any notes or scripts you share, and write down the date, because a model list that changes this fast makes old instructions wrong quickly.

Questions to ask before a team adopts it

The public pages leave several things open, so ask before you commit:

• Data. What happens to prompts and code on each plan? The Business plan runs under Google Cloud terms, and the individual plans point to Google's Privacy Policy.
• Limits. What are the weekly and five-hour limits on the plan you will buy, and are they shared across IDE, CLI and app?
• Billing. Business is listed from $30 per seat per month or pay-as-you-go with a $0 seat fee. Which suits your usage pattern, and who sees the spend caps?
• Source. The CLI installs from a script, so who reviews what it installs on your machines? A security team will want to read the script and pin the version, which the download page shows (1.4.10 when we looked).
• Lock-in. Skills and MCP follow Antigravity's own folders. Keep the substance in AGENTS.md and in the repository, so a move later is cheap.

If you cannot get answers, run the free tier on non-sensitive code first and decide with evidence from your own use.

Which surface should you start with?

The four surfaces answer different habits, so pick by how you already work rather than by feature count.

Nothing stops you from using two. A common pattern is the IDE for reading and editing and the CLI for a long background job. Check that your plan's limits are shared between surfaces before you rely on that.

Antigravity compared with Gemini CLI

Since many readers arrive from Gemini CLI, here is the short version.

The last row matters to some teams. Gemini CLI can be read, forked and pinned. Antigravity CLI is distributed by an install script, and we did not find a source repository on the pages we read.

MCP, skills and plugins

From the migration page you can infer the layout. MCP servers live in mcp_config.json (global or per workspace), skills live in skills folders, and what Gemini called extensions are now plugins. The CLI page confirms slash commands for plugins, MCP, skills and hooks. Specific install commands for an MCP server are not on the pages we read.

Because AGENTS.md is still read, a repository that already serves Claude Code or Codex needs no new rules file. Add the same MCP servers you rely on elsewhere, such as Playwright MCP and GitHub MCP, and see the wider MCP server list.

Limits, privacy and data handling

The homepage makes a general statement: "Google Antigravity is built for user trust". That is not a data policy. The page links to Google's Privacy Policy, and the Business plan runs under Google Cloud Terms of Service.

Until that is answered, treat the free tier the way you treat any free hosted tool: do not point it at code you are not allowed to send to a third party. For work code, the Business plan's Cloud terms and admin controls are the documented path.

Other limits to expect:

• Weekly rate limits on the free plan, with no published numbers.
• Partner-model removals scheduled for 2 November 2026.
• Differences from Gemini CLI that Google itself calls out.

Who it suits

• Individuals who want a free or low-cost agent and an IDE in one place.
• Former Gemini CLI users, who are being pointed here.
• People who want Claude models on a Google plan.
• Teams on Google Cloud that want one invoice and spend caps.

It suits less well if you need a published feature list for hooks, skills and subagents before you commit; the public pages say less on those than Claude Code's docs do.

Alternatives and duos

The whole field is on AI coding agents.

Our take

This comes from Google's public pages, not from a test run.

Antigravity's strongest card is the price of getting in: $0, with the model list showing that even a free account reaches a current Flash model. Add the paid-plan Claude access and a CLI that inherits Gemini CLI's skills and hooks, and it is a serious option for people who do not want to start with a subscription from an AI lab.

The weak spots are information. Prices for the paid Google AI tiers are not on the pricing page, rate limits are not numbers, and the privacy terms for individuals are not spelled out where we looked. We would want those answered before moving a team onto it.

A reasonable path: install the CLI on a personal project, see which model you get, and read the privacy policy before you point it at anything private.

Checklist

• Choose a surface: IDE, CLI or app
• Check which models your plan unlocks in the selector
• Note the 2 November 2026 removal of the Claude 4.6 and GPT-OSS models
• If you came from Gemini CLI, move skills to .agents/skills/ and MCP to mcp_config.json
• Read the privacy policy before using it on private code
• Re-check plan prices, which change often