Google Releases Antigravity, An "Agent-First" Coding Tool
Antigravity is a VS Code fork that is powered by the latest coding models like Gemini 3 Pro, Claude 4.5 Sonnet, and GPT-OSS 120B.
Google recently released Antigravity, a new AI-powered “vibe coding” platform built on top of VS Code. My initial reaction was quite skeptical: Why build yet another one?
Google has a somewhat chaotic product history in this space. In the past five years alone, they have built and launched several AI-powered coding tools. There is Firebase Studio, which is already a full-blown IDE. There is Google AI Studio, which serves as a browser-based app generator powered by Gemini. And then there is Jules, which functions as an AI coding agent.
So, why exactly did Google build Antigravity, and how is it fundamentally different from the others?
Let’s talk about it.
What is Antigravity?
First of all, I have to say that the name “Antigravity” for an IDE sounds terrible. I thought Google had improved at naming its products after the release of Nano Banana, but unfortunately, it turns out they haven’t.
Anyway, Antigravity is what they call an “agentic development platform.” Technically, it is basically just a Visual Studio Code fork with the extended capability of integrating coding models like Gemini, Claude, and GPT.
But Google positions it as more than just a code editor. They call it a “home base” for software development in the era of agents. They want to enable anyone with an idea to build that idea into reality. It is currently in public preview and available for free.
So what makes Antigravity unique?
Google claims this product brings together four key things: Trust, autonomy, feedback, and self-improvement.
Trust. Most AI coding tools today are on two extremes. They either hide everything the agent does or dump a mess of raw code and tool calls on you. Antigravity tries to find a balance. It shows you “Artifacts” like task lists, implementation plans, and screenshots. This way, you can actually verify the work without having to dig through raw logs.
Autonomy. This is the part that interested me the most. You get two different views. The first is the standard “Editor view,” which is your typical IDE experience. The second is the “Manager view.” This is like a mission control where you can spawn agents to do background research or tasks in a different workspace while you focus on something else.
Feedback. Since AI isn’t perfect yet, you need a way to correct it when it messes up. Antigravity lets you give feedback directly on the artifacts it creates. It works a lot like Google Docs comments. You can leave comments on text plans or even select parts of a screenshot to give feedback. The agent then incorporates that feedback without stopping its process.
Self-Improvement. The system treats learning as a core feature. The agent can retrieve information from and contribute to a knowledge base. This means it can save useful code snippets or remember the steps it took to solve a specific problem, so it doesn’t have to figure it out from scratch next time.
You can learn more about these features on the official blog post.
How to get started with Antigravity
Head over to the product page and download the installer. The desktop application currently supports macOS, Windows, and Linux platforms.
Minimum Requirements:
MacOS: Versions with Apple security update support. This is typically the current and two previous versions. Min Version 12 (Monterey), X86 is not supported.
Windows: Windows 10 (64-bit)
Linux: glibc >= 2.28, glibcxx >= 3.4.25 (e.g. Ubuntu 20. Debian 10, Fedora 36, RHEL 8)
I am currently using a MacBook Pro with an M3 chip, so all screenshots in this post are from macOS. Once the app is installed locally, open it and follow the on-screen instructions to get started with the setup process.
I really like how much customization you can do with this IDE right from the start. For example, during installation, you can choose from four different configurations based on how you want to use it:
Secure Mode: This provides enhanced security controls for the Agent, allowing you to restrict its access to external resources and sensitive operations.
Review-driven development: Before executing critical tasks like CLI or script executions, it will ask for your permission first. This is the one that’s recommended by Google and probably what you should be choosing too.
Agent-driven development: If you want to go YOLO and allow the AI agent to execute scripts and CLI commands without your permission, then choose this option.
Custom: You can configure which policies to allow and not allow.
After the configuration is complete, it will ask you to sign in with your Google account. No, you cannot use any other sign-in method like Microsoft or GitHub; you must sign in with a Google account. That is a pretty clever ecosystem move by them.
This is what the initial dashboard looks like:
Planning (Default): The agent takes time to think, organize tasks, and create detailed plans before executing.
Fast: Best for quick fixes like renaming variables or running scripts. The agent skips the planning phase and just does the work immediately.
What’s new to me, though, is the “Agent Manager” feature. When you click on it, it opens a separate window that looks like this:
Google seems to think this is the future. They believe that as models get better, we will stop writing code line-by-line and start managing “work” at this higher level. For now, it is just a toggle away, but they plan to make this the central experience of the platform soon.
Let’s talk about the models
Okay, now let’s take a look at the supported coding models.
Of course, Google is going to prioritize and promote its own Gemini models. You have the Gemini 3.0 Pro, which comes in two versions (High and Low). You also get access to Anthropic’s Claude Sonnet 4.5 and Claude Opus 4.5, both of which include Thinking mode support. There is also GPT-OSS 120B, which is the latest coding model from OpenAI.
Another cool feature that Antigravity has — and which I think solves a lot of my issues with other tools like Cursor or Warp — is the native browser support.
For example, one of my biggest frustrations when “vibe coding” is the constant copying and pasting of console logs between the browser and the IDE. If anything fails, I usually have to add debug logs to the code, copy the logs from the browser, paste them into the IDE, and ask the AI agent to solve the problem for me.
With native browser support, the AI agent has immediate access to the browser, so there is no need to copy and paste anything. It is a huge time saver!
So here’s how it works in Antigravity. First, click on the tiny browser icon in the upper right corner of the IDE, and it will open a browser with the following welcome screen:
Before I end this post, I wanted to show you an example web app created with this tool. In my setup, I chose Gemini 3 Pro in Planning mode. I wanted to create a simple image generator web application powered by Nano Banana Pro.
Here are my instructions:
Prompt: Help me create a simple web application that accepts an image input of any product and transform them into a studio quality image. Use Nano Banana Pro image model to transform the product image. The user interface will be a two column layout with the left panel for the input image and the output image on the right panel. Add a “generate”, “regenerate”, and “clear” buttons at the bottom of the dashboard. Use a square aspect ratio in the output image. Make sure the app only accepts image files and is responsive.
The first thing it did was compose a detailed set of tasks on how it plans to build the app.
You can review the plan and click on the Proceed button to approve it and let the AI agent continue with the coding phase. In case you want to make changes — suppose you wanted to change the framework from Vite to NextJS — you can simply instruct it in the prompt box.
This is what the UI looks like:
Okay, let’s try to see if the image generation works. I uploaded a sample image of a headphone and asked it to improve it for a product campaign poster. Here’s the prompt used by the web app:
Prompt: Transform this product image into a high-quality studio professional product photography. Shot in a different angle to focus on the product. Add text for a promotional poster. Maintain the product details exactly but enhance the lighting, angle, zoom level and presentation.
Perfect! I love how Nano Banana Pro handled the product image here. The text rendering is clean and the shadows are realistically added to the surface. Of course, you can play with the prompt to adjust the output image. The prompt is located in the route file under the API/generate folder.
import { GoogleGenerativeAI } from “@google/generative-ai”;
import { NextResponse } from “next/server”;
// Hardcoded for this demo environment where .env creation is restricted
const API_KEY = process.env.GOOGLE_API_KEY || “xxxx”;
const genAI = new GoogleGenerativeAI(API_KEY);
// export const runtime = “edge”; // Reverting to default Node.js runtime for better compatibility
export async function POST(req: Request) {
try {
const formData = await req.formData();
const imageFile = formData.get(”image”) as File;
const prompt = formData.get(”prompt”) as string || “Transform this product image into a high-quality studio professional product photography. Shot in a different angle to focus on the product. Add text for a promotional poster. Maintain the product details exactly but enhance the lighting, angle, zoom level and presentation.”;
if (!imageFile) {
return NextResponse.json({ error: “No image provided” }, { status: 400 });
}
const arrayBuffer = await imageFile.arrayBuffer();
const base64Image = Buffer.from(arrayBuffer).toString(”base64”);
// Initialize the model
// Note: ‘gemini-3-pro-image-preview’ is the researched ID for Nano Banana Pro.
// Testing with ‘gemini-1.5-pro’ as a fallback if the ID fails, but keeping original for now.
const model = genAI.getGenerativeModel({ model: “models/gemini-3-pro-image-preview” });
const imagePart = {
inlineData: {
data: base64Image,
mimeType: imageFile.type,
},
};
const result = await model.generateContent([
prompt,
imagePart
]);
const response = await result.response;
// Log success for debugging
console.log(”Gemini API Response:”, JSON.stringify(response, null, 2));
return NextResponse.json({ result: response });
} catch (error: any) {
console.error(”API Error Detailed:”, error);
// Return the full error message to the client
return NextResponse.json({
error: error.message || “An unknown error occurred”,
details: JSON.stringify(error, Object.getOwnPropertyNames(error))
}, { status: 500 });
}
}Final Thoughts
For now, Antigravity is yet another vibe coding tool. It is nothing too special just yet. There’s no reason for you to panic and cancel your Cursor subscription to make the switch immediately.
But in the longer term, Antigravity poses a huge threat to the competitors. Seeing how aggressive Google is in pushing AI tools and services, and knowing how much money and talent resources they have, I can definitely see it becoming a go-to vibe coding tool for most people.
I love the level of autonomy here. In the example I gave, I used a simple instruction to use the Nano Banana image model without a specific context, like a sample code snippet, but it was able to do it perfectly. It tries to search the internet to find appropriate documentation on its own.
Another thing that impressed me is the fact that the output works as expected. In less than 15 minutes, the tool was able to give me a working application. The prompt I gave it was somewhat vague, but the output was very close to what I had in mind.
Do I recommend making the switch to Antigravity? For now, I encourage you to go try it yourself. Google gives you a generous number of free credits to explore and try building simple web apps. Have fun!
Hi there! Thanks for making it to the end of this post! My name is Jim, and I’m an AI enthusiast passionate about exploring the latest news, guides, and insights in the world of generative AI. If you’ve enjoyed this content and would like to support my work, consider becoming a paid subscriber. Your support means a lot!
human slop