A few months ago, "an AI agent in your terminal" meant one or two tools that early adopters argued about. As of this week, all three of the biggest AI labs have one.

On 25 June 2025, Google released Gemini CLI, an open-source AI agent that runs in your terminal. It joins Anthropic's Claude Code and OpenAI's Codex CLI. For developers, that's good news, and also a new question: which one should you actually use?

Here's what was announced, why it matters, and how I'd think about choosing.

What Google released

The short version:

• A terminal agent. You run it inside a project folder and give it tasks in plain language. Like the others, it can read your code, propose and make changes, and run commands, rather than just answering questions in a chat window.
• Open source under Apache 2.0. The code is on GitHub. You can read how it works, file issues, and contribute.
• A generous free tier. Signing in with a personal Google account gives you free usage, which lowers the barrier to trying an agent to almost nothing.
• Powered by Gemini. It uses Google's Gemini models, including Gemini 2.5 Pro, which has had a strong reputation for coding since its release in March.

If you have Node.js installed, trying it is one command:

npm install -g @google/gemini-cli
cd your-project
gemini

Why this is a bigger deal than one more tool

When one company ships a new kind of product, it could be an experiment. When the three largest AI labs all ship the same kind of product within a few months, it's a direction.

The terminal agent is becoming a standard way of working with AI on code, sitting alongside editor-based assistants like GitHub Copilot and Cursor rather than replacing them. The agent works where your code, Git, tests and build tools already live. It can run your test suite, see the failure and try again. That loop is what makes agents feel different from autocomplete.

Competition also tends to be good for the people using the tools. Free tiers, open source code and fast iteration all push the whole category forward.

The three, side by side

I use Claude Code every day, so I'm not neutral. But the honest answer is that all three are capable, and the best one depends more on your situation than on any feature list.

How to choose

These are the questions I'd ask, in roughly this order:

1. Which model works best on your code?

The agent is mostly a wrapper; the model does the thinking. Models differ by language, framework and task type. Try each one on the same real task from your backlog, something with a few files and a test to pass. One afternoon of that is worth more than any review or leaderboard.

2. How do you want to pay?

A free tier is perfect for learning and side projects. For daily professional use, think about predictability: a subscription, API usage, or a company account. Agents can use a lot of tokens on big tasks.

3. Does open source matter to you?

If you want to audit what the tool does, extend it, or run it in a tightly controlled environment, Gemini CLI and Codex CLI being open source is a real advantage.

4. What are the data terms?

Before pointing any agent at a work repository, read how your prompts and code are handled, especially on free tiers, and check what your company allows. This is the step people skip and shouldn't.

5. What does your team already use?

Shared instruction files, shared conventions and shared experience matter. A team that knows one tool well usually gets more from it than from switching every month.

What I'd do this week

If you've never used a terminal agent, Gemini CLI's free tier makes this the easiest moment yet to try one. Install it on a side project, ask it to explain the codebase, then give it one small, real task with a test.

If you already use Claude Code or Codex CLI, give Gemini CLI the same task you did last week and compare. Not to switch for the sake of it, but to know what each one is good at.

And whichever you use, keep the same rule: the agent writes the code, you own it.

Have you tried more than one terminal agent yet? Which one fits your workflow, and why?