First Look: Gemini CLI vs Claude Code

I'm using this over Claude Code to save 💵 next month

Alright, we’re going to go through this pretty quick as it just released yesterday morning. First, head over and download the Gemini CLI if you want to follow along.

First things first, I opened the CLI and had to log in with Google. I tried to log in with a Google Workspace account, but it wouldn't let me unless I pay and enable the Gemini API….

Thankfully, Google is offering Gemini free for personal Gmail accounts. I just logged into my regular gmail and was able to get right into it!

In the official blog post, they're announcing a pretty good free tier. I’m excited to see how much I can use this over the following month:

To ensure you rarely, if ever, hit a limit during this preview, we offer the industry’s largest allowance: 60 model requests per minute and 1,000 requests per day at no charge.

What’s the next step after I log into Gemini? I have to add our issue tracker at work, obviously 🙂

Adding our first MCP Server

It’s almost the same as Claude Code. As soon as I opened it, I typed /mcp as you can see below:

The first time you run it and try to do anything, it kicks you out to this webpage.

It describes the use of the `mcpServers` object key, just like every other MCP tool has been using, including cursor, cline, claude, etc.

For now there isn’t a fancy CLI arg to add servers that I can see, so here’s how to do it:

vi ~/.gemini/settings.json 

{
  "theme": "Default",
  "selectedAuthType": "oauth-personal",
  "mcpServers": {
    "linear": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.linear.app/sse"]
    }
  }

I decided to use the MCP Remote proxy since it works perfectly with OAuth for now, as I’m unsure of the exact support Gemini has internally.

As soon as that’s done, I rerun the `gemini` command, and here's what we get after running /mcp inside of it:

Summary

I tried to use it for a few tasks today. It seems to use tool calls perfectly as I would expect, but agent tool loop needs some work.

Oddly enough, I asked it to perform a task, and in the middle it just stopped. I had to send “ok keep going,” to get it to finish working on something I gave it. This happened once and I’m not sure how often it will do that as I haven’t had enough test time.

I also had to hand hold it a lot more than claude, which figures out the correct approach faster. However, you can’t beat free vs $100 monthly right now, so I will keep trying it out.

I’m excited because I was one day away from renewing Claude Max at $100 a month, too. Now, I’ll be using it for free during the preview period, and I hope it lasts for quite a while.

I yearn for the day we we can do the same stuff locally… Thankfully, Google's Gemini CLI is open source, so I'm hoping we can either fork or add a feature to it to select custom AI providers soon. It's a pretty solid CLI tool, it definitely feels on par with Claude Code in terms of the user interface, the speed, and the interactions that take place while working with it.