Antigravity (currently Antigravity 2.0) is Google's agent-first AI coding platform and the successor to the deprecated Gemini CLI. It is no longer just an IDE plugin, but a product family with four surfaces:
- Antigravity Desktop App: a standalone agentic development workbench
- Antigravity IDE: an agent-first IDE with Agent Manager, Artifacts, and deep codebase understanding
- Antigravity CLI (
agy): a lightweight, fast, terminal-native agent - Antigravity SDK: a Python toolkit for prototyping custom agents and evaluations
Antigravity defaults to Gemini 3.5 / 3.6 / 3.7 Flash models and also supports switching to Claude, GPT-OSS, and other models via a model picker.
Antigravity treats the human as a manager of agents:
- The unit of work is a task or a plan, not a keystroke
- Agents have direct access to the editor, terminal, and an integrated browser
- Every run produces Artifacts — task lists, plans, screenshots, browser recordings — so the human can verify what was done before accepting it
Whether you use the desktop app, IDE, terminal CLI, or SDK, the same Antigravity agent engine powers Projects, multi-workspace orchestration, subagents, and scheduled tasks.
- Default models: Gemini 3.5 Flash / 3.6 Flash / 3.7 Flash
- Optional models: Gemini 3 Pro, Claude family, GPT-OSS-120B
- Switch quickly with
/modelor configuration
- Artifacts make agent runs auditable
- Agents accumulate learnings across sessions instead of starting cold every time
| Surface | Positioning | Best For |
|---|---|---|
| Desktop App | Standalone agentic workbench | Daily development, multi-project management |
| IDE | Agent-first IDE with Agent Manager | Deep code editing and parallel agents |
CLI (agy) |
Terminal-native agent | Command-line users, scripting / CI |
| SDK | Python programmable interface | Custom agents and evaluations |
- Multi-step task execution: plan → edit → run → verify
- Parallel agents working across files and even repositories
- Long-horizon refactors and migrations rather than single-snippet edits
- Agent Manager: fan out and orchestrate tasks across multiple agents and workspaces
- Projects: group related conversations and tasks
- Subagents: delegate focused work to specialized agents
- Agents can drive a real browser to test the changes they wrote
- Captures screenshots and recordings as evidence of behavior
- Carries project context across agent runs
- Learns from prior interactions, including corrections you apply
- Supports Hooks and scheduled tasks
Antigravity is currently in public preview and free for individual users. Download the desktop app or IDE for your system from:
System requirements:
- Windows 10 (64-bit) or later
- macOS Monterey (12) or later
- 64-bit Linux (modern glibc / glibcxx)
- Launch Antigravity and sign in with your Google account
- Pick a default model — Gemini 3.5 Flash is a good baseline
- Open a folder/workspace and start in Editor/IDE view, or jump to Manager view to plan a multi-agent run
- Describe the task in natural language ("add unit tests for
auth/", "migrate this service from REST to gRPC") - The agent produces a plan as an Artifact for you to approve
- The agent edits files, runs commands in the terminal, and optionally drives the browser to verify
- Review the diff and the Artifacts (screenshots, recordings, logs), then accept or iterate
Antigravity 2.0's terminal entry point is agy (Antigravity CLI), the successor to Gemini CLI.
- Via IDE / Desktop App: open the Command Palette (F1 / Ctrl+Shift+P) and run
Shell Command: Install 'agy' command in PATH - Standalone install (see official docs):
# Follow the official Antigravity documentation for the latest command
# Launch interactive terminal session
agy
# Ask a prompt directly
agy "Analyze current project structure"
# Non-interactive (headless / script) mode
agy -p "Write unit tests for src/auth.js"
# Migrate plugins from Gemini CLI
agy plugin import gemini-m, --model <model>: specify model, e.g.gemini-3.5-flash-p, --prompt <prompt>: run in non-interactive mode--approval-mode <mode>: approval mode (default,auto_edit,yolo,plan)-s, --sandbox: run inside a sandboxagy mcp list/agy skills list/agy hooks list: manage MCP, Skills, Hooks
Note: the
geminicommand stopped serving individual users on June 18, 2026. Migrate scripts toagy.
Hand the agent a multi-file refactor and review the produced plan and diff instead of writing every edit yourself.
Spec → implementation → tests → browser-verified UI changes, all driven by one or more agents in the same workspace.
Use Manager view to run several agents on different tasks/branches at once and inspect their Artifacts as they finish.
Use the Antigravity SDK to quickly prototype custom agents in Python and run evaluations.
- Start from a clear task statement — the better the prompt, the better the plan
- Review the Artifact, not just the diff — screenshots and recordings catch behavior the diff hides
- Let the agent verify — give it permission to run tests and drive the browser; the goal is verifiable output
- Bring corrections back as context — Antigravity learns across runs, so explicit feedback compounds
- Start with the desktop app or IDE — move to the
agyCLI for scripting once you're comfortable
Antigravity is a proprietary Google product. It is free for individual users during the public preview; some advanced features or higher quotas may require a Google AI subscription. Subject to Google's terms of service.