← Back to Course|Developer Tools
🪐
Agent-First IDEFree Public Preview

Google Antigravity

Google's agent-first development platform - autonomous agents that plan, build, and test across your whole project

Announced at Google I/O 2026, Antigravity is Google's standalone, agent-first development platform. Instead of autocompleting line by line, you describe a task in plain language and autonomous agents plan it, edit across multiple files, and actually run and test the app before handing it back. It is built on Gemini 3.5, and a signature feature is the Browser Subagent: Antigravity spins up a real Chrome window, clicks through your app, takes screenshots, and fixes what is broken - without you ever opening DevTools.

Made by
Google
Type
Agent-first IDE
Free tier
Public preview
Powered by
Gemini 3.5
Platform
Mac, Win, Linux
Get it
antigravity.google
Where it fits in Google's tools:

Google's old browser IDE, Project IDX, became Firebase Studio(April 2025), which is now winding down - Google points users toward Google AI Studio(for prompting/prototyping with Gemini) and Antigravity (the agent-first IDE for building real apps). If you want autonomous agents writing and testing code, Antigravity is the current tool.

🪐 Install & Setup

Step 1 - Download the app

Go to antigravity.google and download the desktop app for Mac, Windows, or Linux. It is a standalone app (free public preview) - it does not replace your existing editor. There is also an Antigravity CLI and a mobile companion app in preview.

Step 2 - Sign in with Google

Log in with your Google account. The preview is free for individuals; usage limits scale with your Google AI plan (AI Pro, or AI Ultra for heavy use).

Step 3 - Open or create a project

Point Antigravity at a folder (existing repo or a new one). The agents read your whole project as context before they start.

Step 4 - Connect tools (optional)

Antigravity supports MCP servers, reusable Skills, and JSON hooks (the same building blocks you learned in Levels 7-8). Add the ones your workflow needs - e.g. GitHub, a database, or your issue tracker.

🔑 Key Features

🧑‍✈️
Agent Manager
A mission-control view where you launch agents, watch their plans, and approve or redirect them. You manage agents, not keystrokes.
🌐
Browser Subagent
Spins up a real Chrome instance, uses your app like a person - clicks, fills forms, screenshots - and self-debugs anything that breaks.
⚙️
Parallel Subagents
Splits a big task across multiple agents working at once, each in its own clean context (the Level 8 idea, built in).
📦
Artifacts
Agents produce reviewable artifacts - task lists, plans, screenshots, test results - so you can verify the work, not just the final diff.
🔌
MCP + Skills + Hooks
Connect external tools via MCP, encode know-how as Skills, and trigger shell commands with JSON hooks. Scheduled tasks too.
⌨️
CLI & SDK
Drive the same agents from your terminal or from your own code, and run them on managed cloud infrastructure.

🚀 Example Project: Build a To-Do App

Here is the full agent-first loop on a small but real project - the same shape you would use for a feature at work.

1. Describe the app in one prompt

In the Agent Manager, give a single high-level task:

"Build a to-do web app: add tasks, mark them done, delete them, and save them in the browser so they persist on refresh. Plain HTML/CSS/JS, no backend. Make it keyboard accessible."
2. Review the agent's plan (artifact)

Antigravity returns a plan before writing code. You approve or tweak it:

Plan 1. index.html - input + add button + task list 2. styles.css - clean layout, focus states for a11y 3. app.js - add / toggle / delete, render from state 4. persistence - save state to localStorage, load on start 5. verify in browser: add 3 tasks, toggle, delete, refresh
3. Let it build - then watch it test itself

The agent writes the files, then the Browser Subagent opens the app and actually uses it: types a task, clicks add, toggles it done, deletes one, refreshes the page to confirm it persists. It captures screenshots at each step. If "delete" does not work, it sees the failure, fixes app.js, and re-tests - on its own.

4. Review artifacts & iterate

You get the working app plus screenshots and a test summary. Send one follow-up to refine:

"Add a filter: All / Active / Done. Keep the current task in focus after toggling. Re-test in the browser."

💡 Pro Tips

  • Review the artifacts, not just the result - read the plan and the screenshots before approving. Agent-first does not mean hands-off (Level 7: stay the engineer).
  • Think in tasks, not keystrokes - write a clear goal and acceptance criteria, the way you learned in Level 9 (spec-driven development).
  • Let the Browser Subagent earn its keep - ask it to verify the actual user flow, not just "does it compile."
  • Connect MCP servers for your real tools (GitHub, database, tracker) so agents can act, not just suggest.
  • It is a free preview - usage is capped by your Google AI plan tier, so keep tasks scoped.

⚖️ Which Coding Tool Should I Pick?

The current tool is highlighted. There is no single winner - pick by how you like to work.

ToolTypePricingBest forModels / AI
GitHub CopilotIDE extensionFree tier; Pro ~$10/moAI in your existing editorGPT-5.5, Claude Sonnet 4.6, Gemini 3.5
CursorAI IDE (VS Code fork)Hobby free; Pro $20/moAll-round AI editor + agentGPT-5.5, Claude Opus 4.8 / Sonnet 4.6
WindsurfAI IDE (VS Code fork)Free tier; paid plansAgentic flows (Cascade)Frontier models + own
Google Antigravity(this page)Agent-first IDEFree public previewAutonomous build + browser testGemini 3.5
Claude Code (CLI)Terminal agentClaude sub or APIAgentic coding in the terminalClaude Opus 4.8 / Sonnet 4.6 / Haiku 4.5
VS Code + AI extEditor + extensionsFree options existKeep VS Code, add AIDepends on extension

Pricing and model support change often. Verified June 2026 - confirm on each tool's official site.

🛠️ Troubleshooting & Common Pitfalls

Can't sign in or no preview access
Antigravity is a public preview - sign in with a Google account. If access is gated, request it on antigravity.google and check your Google AI plan.
The Browser Subagent does nothing
It needs a Chrome install it can drive. Make sure Chrome is installed and permitted, then re-run the task so the agent can open and test the app.
The agent goes off track
Open its artifact (the plan), correct it, and re-run. Agent-first still needs your review at each gate - do not just approve blindly.
Hitting usage limits
Free-preview usage scales with your Google AI tier (Pro / Ultra). Keep tasks scoped, or upgrade the plan.

📚 Related Course Lessons

Go deeper on the skills this tool is for:

L46 - Agentic CodingL56 - Prompter to Loop DesignerL63 - The Minimum Viable Loop

🔗 Official Resources

Get AntigravityDownload the agent-first IDE (public preview)Announcement & overviewGoogle Developers blog: features explainedGoogle AI plansAI Pro / AI Ultra usage limitsGoogle AI StudioPrompt & prototype with GeminiI/O 2026 dev highlightsContext on the 2026 launch

Links open the official sites. Pricing and features change often - always confirm there. (Verified June 2026.)

→ Compare: Cursor→ Compare: Windsurf→ Compare: Claude Code CLI