So, Google just released a new CLI coding agent, and yes, it's open source, completely free, and powered by Gemini 2.5 Pro with an impressive 1M token context window. 🤯
In this post, we'll take a quick dive into what Gemini CLI is all about, how it compares to the best in the market, like Claude Code, where it excels (and also where it struggles!), and my opinion on whether it's worth switching to.
We'll end the post by asking it to build a quick mini-project to get an idea of how good it is.
Let's jump in!
Brief on Gemini CLI
Gemini CLI is Google's free, open-source terminal agent completely written in TypeScript and powered by Gemini 2.5 Pro (a model with a massive 1 million token context window). It is available via Gemini Code Assist (60 requests/min, 1,000/day).
It does support Model Context Protocol (MCP) servers and everything you'd expect, like codebase understanding, code/test generation, bug fixes, and all of that.
The Gemini CLI uses a reason and act (ReAct) loop with your built-in tools and local or remote MCP servers to handle complex tasks like fixing bugs, adding new features, and improving test coverage.
In short, it's Google's response to Anthropic Claude Code.
Here are some of the key capabilities of Gemini CLI:
- Code Editing and Refactoring: It can understand your codebase well and can especially help refactor your codebase more easily.
- Bug Fixing: It goes without saying, since it gets auto context on your codebase, it can help you find and fix bugs.
- Test and Documentation Generation: It can generate tests for your codebase and can also help you write documentation, especially for wikis and READMEs.
- Manage Conversation History and Memory: Similar to how you can manage your chats and memory on ChatGPT and stuff, you can do the same with Gemini CLI. This allows you to keep things separated and get a much better response from the AI model.
All you need to use this tool is Node.js, a personal Google account, and, obviously, a terminal. 😉
The team has shared this cool demo showing how this AI agent can use MCP servers (in this case, Veo and Imagen) to generate a short video of a cat adventure around Australia.
Look, we're talking 60 model requests per minute and 1,000 requests per day at zero cost with all of these features. I think it's a win-win. 😋
You don't pay much and even get all of these features for free, which would otherwise cost around $200/month for Claude Code, so you know how big of a deal this is.
Is this just another Claude Code?
I wouldn't say it is just another Claude Code, but it's definitely just another CLI coding AI Agent.
So what's new about it? To me, the biggest selling point for Gemini CLI is that it's completely free with nicer usage limits and open source compared to the closed source Claude Code.
Anything special other than that? No!
Gemini's recent models have been amazing with coding. Gemini 2.5 Pro is one of the best yet cheapest models right now when it comes to coding, and Gemini CLI brings all of that right to the terminal.
You get a massive 1 million token context window model, Gemini 2.5 Pro, all for free? What else can you even ask for, right?
It seems like the tough AI competitors are fighting for a spot in the terminal, essentially taking on the developer AI ecosystem.
First, there was Claude Code released in February, then OpenAI Codex in April, and now Gemini CLI in June.
So basically, you've got a "terminal agent" from all three AI giants! 🥴
My experience with it compared to Claude Code
So does it compare? I would say straight up NO!
Don't get me wrong, it's great, it has a super cool free usage limit, and you can go test it right now. The fact that it's completely open source with the Apache 2.0 License is great.
But, when compared to Claude Code, it simply does not stand any chance. Anthropic does some real magic when it comes to working with agentic features.
The tool itself is great, but I've run into quite a few issues with it:
- Extremely slow: I'm not exaggerating, it's just a bit too slow. Even the smallest changes take over 5-10 minutes.
- Feels super dumb: In these two or three days of testing this AI agent, it feels like you're not really talking with Gemini 2.5 Pro but a dumb version of it. I've gotten much better results in Google AI Studio for the same task.
- Lots of errors: It's a fairly new tool, so it's likely to have errors. I don't know if it's fair to compare on this just yet, but I've run into a lot of errors where it just crashes (runtime errors!).
You'll see a new issue raised on GitHub every few minutes, and yes, every few minutes!
Quick Coding Test
Enough of the theories, let's see how good and quick it can code a simple ping-pong game. I'm doing this quick test just to give you an idea of how well it works and how long it takes to complete a project.
So many folks are building stuff with Gemini CLI. Take this as a reference: 👇
https://x.com/sawyerhood/status/1912693186474222050
1. Ping-Pong Game
Prompt: Build a cool looking ping pong game in a single file using HTML, CSS and JS. Include dynamic lighting, shadows and smooth paddle/ball animation.
This is a somewhat updated version of the same prompt where I've asked it to add a few new features on top, and here's how it went.
Here's the code it generated: Link
Here's the output of the program:
This one didn't take much time. I was able to wrap it up within 2–3 minutes. It was just a simple one-file code, and I must say, it did a decent job here.
However, it didn't implement the lighting, and the implementation is not completely correct. The ball spawns randomly in the middle of the game at times, and also the shadows and the UI are not that good-looking, but that's fine.
Overall, as you can see, it can do a decent job in coding, and for a decent "vibe coding" session, it can be a great help.
Conclusion
Gemini CLI shines in the generosity of its usage limits and open source. But, you cannot compare it to Claude Code in terms of performance and overall quality (not yet!).
Still, for a tool that gets you Gemini 2.5 Pro right in the terminal with real-time web context, multi-modal outputs, and complete API/script integration, all completely free and open source, it is basically a god's work. This will make it more accessible as a terminal AI agent than ever.
It may not be the best tool right now, but it's definitely a serious player in the terminal AI ecosystem.
It's a no-brainer alternative to Claude Code if you are looking for a cheaper yet similar tool, and I'd definitely suggest you give this a shot.
🤔 And by the way, why is every new coding agent CLI tool using TypeScript? Like, does this make sense? Use Node for CLI when better alternatives like Go or even Rust exist?
