Google Antigravity: A Beginner's Guide While Gemini 3.8 Is Free
Google's agent-first IDE now comes with a free Gemini 3.8 Flash trial — a rare window to learn agentic development at zero cost. This guide walks through installing Antigravity, the Planning workflow and Artifacts, the Browser Subagent, MCP servers and Skills, plus the free-tier caveats to know before you commit.
Antigravity launched in November 2025 as Google's answer to Cursor and Windsurf, went 2.0 at I/O May 2026, and just became the cheapest way to try frontier agentic coding: with Gemini 3.8 Flash's arrival, Antigravity is offering it on a free trial. If you've been waiting for a reason to learn agent-first development, this is the window.
Step 1: Install and Find Your Three Surfaces
Grab the desktop app for macOS, Windows or Linux (or the VS Code extension, which auto-installs the local `agy` backend — no terminal setup needed). Sign in with a Google Account; the free tier is enough to start. You get three surfaces: the Editor (VS Code-style, files on the left), the Agent Panel on the right (`Ctrl+L` to focus, `@` to mention files, `/` for commands), and the Agent Manager (`Ctrl+E`) — a separate view with an Inbox for finished work, Workspaces, a Playground for throwaway sessions, a Knowledge area and a built-in browser.
Step 2: Plan, Review Artifacts, Then Execute
Pick Planning mode for anything non-trivial. The agent first produces Artifacts — a Task list, an Implementation Plan and later a Walkthrough — and waits for your review before touching code. Approve, and it executes in Task Groups: top-level goal, editable file pills, collapsible progress steps. This plan-then-implement loop is the core habit of agentic development: the spec guides the code, not the other way around. Use Fast mode for one-liners. In Settings → Agent, you decide the autonomy level: Request Review or Always Proceed for artifacts, plus terminal command Allow/Deny lists.
Step 3: Let the Agent Verify Its Own Work
The killer feature is the Browser Subagent (`/browser`) — it spins up a real Chrome instance, clicks through your web app, and reports what actually works (set Chrome as your default browser first). Pair it with scheduled tasks (`/schedule` or the Schedule UI) for recurring jobs like a daily build check. For grounding, connect MCP servers in Settings → Customizations → Add MCP+ — Google Cloud integrations are one-click, and remote servers just need entries in `~/.gemini/config/mcp_config.json`. Finally, add Skills: directory packages (a `SKILL.md` plus optional scripts, references and assets) under `~/.gemini/config/skills/` (global) or `.agents/skills/` (per project) that load only when a request matches — keeping context lean.
The Caveats
The free tier's quota has been cut at least four times between December 2025 and March 2026, there's no published SLA, and using Claude models inside Antigravity requires your own Anthropic API key. Treat today's free trial as a learning window, not a production plan. But if you learn the Planning–Artifacts–Browser-verification loop now, you'll carry it to any agent IDE you use next.