I’ll admit it, I’m a bit of an AI-sceptic. I see how AI-powered search can perform wonders with natural-language queries, but when it comes to editing my code, I still like to have control.
However, Antigravity could be the tool to convince me otherwise, and if you’re still on the fence about AI coding agents, it may do the same for you, too. This is a tale of false assumptions, one AI pitted against another, and a minor redemption arc.
What is Antigravity CLI?
AI on the command line, with a great-looking app
The Antigravity CLI is a replacement for Gemini CLI, which made the Gemini 3 model available on the command line. It’s an interactive program that lets you query an AI agent for all manner of coding tasks:
Antigravity CLI supports the Gemini 3.5 Flash and Gemini 3.1 Pro models, alongside others like Claude Sonnet, Claude Opus, and GPT-OSS 120B.
When you run it and grant it access to your current directory, it will analyze the files inside, gaining deep insight about the code. Antigravity can make code edits if you allow it, but it can also help you understand a codebase and even just explain how to use a command.
Installing and using agy
A bit of one-off configuration, then usage is straightforward
You can find installation instructions for your platform on Antigravity’s download page. For Linux and macOS users, it’s a simple case of running this command:
curl -fsSL https://antigravity.google/cli/install.sh | bash
Once you’ve installed it, make sure the binary location is in your PATH, then run agy and you should see an initial welcome screen.
Sign in with your Google credentials, and complete the authentication process to continue. You’ll typically see a confirmation screen for new projects, asking you if you trust the contents of whichever directory you launched agy from.
Be aware that Antigravity can change your code and run commands on your machine. Go slowly and pay attention to the on-screen warnings!
Exploring a codebase with agy
Run it from a repo and get all the answers you need
I initially turned to Antigravity to resolve a tricky problem with eza, the modern alternative to ls. Trying to understand the tool’s –total-size option, I began with a simple DuckDuckGo query that relayed some answers using its AI summary feature:
Although AI-powered search can often be wrong, I’ve found it useful most of the time, and it’s certainly less controversial than generative AI, partly because old-school search was already doing much of the same thing anyway.
Something still wasn’t quite right, however: eza was reporting a size of 7,808B while the find command gave 7,637B. At this point, if I’d thought about the problem some more, I may have come to a realization of my own. But I wanted to give agy a run anyway, so I went ahead and fired it up:
Note that I’m running agy directly from the eza source directory here, so there’s no need to give any context like which tool I’m even talking about. Antigravity is clever enough to realize that I’m talking about the eza command and its –total-size option.
While the model thinks, it shows fascinating insights into the process:
As it goes about the job of answering your question, you’ll see references to files, details about the commands it runs, and more. Above all else, being able to peek behind the curtain to see what the agent is actually doing was a revelation for me.
Initially, though, while agy offered amazing insight into some of the possible factors, I wasn’t all too sure they applied, so I went for a follow-up:
By this point, I really wanted to get to the heart of the issue: why find was reporting a different size than eza was. Giving the model additional information can, of course, improve its handling of the task, especially external information such as this.
Sure enough, after a few more seconds, agy came back with plenty of possible explanations.
Using agy—or any other AI tool—can involve refinement and a back-and-forth process. You can always use it in parallel, while you explore the codebase in another terminal or a text editor.
Now, to really test Antigravity, I wanted to see if it could fix my find command to resolve the inconsistency:
In particular, I’m asking it to match eza. There’s not a lot of detail there, but in the context of the overall conversation, agy has no problem at all understanding what I’m getting at.
But there’s still a minor problem: macOS’s BSD version of stat behaves differently from the GNU version of stat. By this point, I’d gotten a bit lazy, so I just banked on Antigravity to do all the work for me:
After just a few minutes, Antigravity had delved into a codebase, explored the path of a specific command-line option, and helped me refine a command for my platform. That final command delivered, now matching the output from eza:
Antigravity has convinced me: an AI assistant can be valuable
Even though I approached it with skepticism, Antigravity delivered. As an affirmed lover of the command line, agy is a powerful yet accessible tool that I’ll be keeping a close eye on in the future.
Stephan is the sports journalist for the Maple Grove Report.
