Lately, I've been using AI agents and I've found them pretty useful. After spending a month with Claude Code and having a great experience, I was excited to try Google's Gemini CLI yesterday. I was really disappointed.

It's not about the quality of the code, the responses, or how the model performs. I never got that far because I couldn't make it work. It's about the experience it offers - or lack of it. In such a competitive environment, it's crazy to me that I couldn't even install and run Gemini.

Authenticate or not

After installing it with npm, I got the first problem. Gemini CLI didn't work properly on my development machine because I don't have a browser there, and the authentication flow wasn't ready for that scenario.

After digging in GitHub issues, I found that the --debug shows the authentication URL. Then I had to:

  1. Copy the URL manually
  2. Set up SSH port forwarding for the OAuth callback
  3. Complete the authentication in my Macbook

Do you have a project?

Once I finally got past the authentication, I thought I could write my first prompt. Wrong. Gemini CLI asked me to create a Google Cloud Project because I was using a Workspace account. Fair enough, let's create it.

After creating the project, I tried to run Gemini CLI again. Another error: I needed to set GOOGLE_PROJECT_ID environment variable before running the CLI. Why couldn't it just ask me which project to use from the ones available in my account?

"Free Tier"

Just when I thought I could finally use the tool, I hit a wall. Google wanted me to enable a paid Gemini API in my Google Cloud Project. Here's where it gets really frustrating as I couldn't test it in advance and the pricing is not that easy to follow.

At this point, I gave up. I've been using Claude Code for a month without any of these issues. It just works. Download, authenticate, and start coding.

Why This Matters

In the AI tools space, the competition is huge. When developers want to try a new tool, especially one that competes with something they're already happy with, the onboarding experience is critical.

Google has great AI capabilities (Veo 3 is so cool!), but if developers can't easily access them, what's the point? With Claude Code, I was productive within minutes. With Gemini CLI, I couldn't run a prompt.