Claude Code vs Gemini CLI vs Codex on a Mac

Claude Code, OpenAI’s Codex CLI and Google’s Gemini CLI are the three terminal coding agents most people are actually choosing between. They look nearly identical from the outside — you type a request, they read your repo and do work — and they feel quite different after a week.

This is a comparison of temperament rather than benchmarks. Benchmarks for these tools go stale in about six weeks, and so does any pricing table, so there isn’t one here: check the vendor’s current page before you commit money.

Claude Code

The implementer. Give it a goal that spans several files and it will plan, read what it needs, and land the change with the least hand-holding of the three. It is the one that most often does the thing you meant rather than the thing you said.

Its other distinguishing property is the one people complain about first and miss most when it is gone: it asks before it acts. Before an edit, a command, a push, it stops and waits for a yes. You can allow once, allow for the session, or refuse. That makes it the most comfortable of the three to point at a repository you care about.

It needs a paid Claude plan. There is no meaningful free tier.

Codex CLI

The reviewer. Codex is strongest reading code and reasoning about whether a change is right, and it is the most at home in a pipeline — GitHub flow, pull requests, CI. If your use case is “look at this diff and tell me what is wrong with it” rather than “write this feature”, it is a genuinely different experience from Claude Code and often a better one.

Access typically comes with a ChatGPT plan, with a per-token API option for automation. If you already pay OpenAI for something else, trying it costs nothing extra.

Gemini CLI

The one with the enormous context window. The headline property is that it can hold far more of a repository in mind at once, which matters for a specific job: understanding an unfamiliar, sprawling codebase. “Explain how this 400-file service fits together” is a question it answers better than the others.

Its access story changed in 2026. Since 18 June, Gemini CLI no longer serves the free tier or Google AI Pro and Ultra accounts. Those users moved to Google’s Antigravity CLI (agy). Gemini CLI is still Apache-2.0 and still works with a paid Gemini API key or a company Code Assist licence. Gemini CLI stopped working? Moving to Antigravity on Mac covers who is affected and how to move your setup.

In practice: use more than one

The sensible answer, if you can, is not to pick.

• Gemini CLI to understand a codebase you did not write.
• Claude Code to make the change.
• Codex to review the diff before it goes anywhere.

That sounds like overhead and mostly is not: they all operate on the same working directory, and switching is a matter of running a different command in the same terminal.

What it does create is a practical problem. Three long-running agents, three terminal windows, and no ambient sense of which one is working, which one finished, and which one has been waiting eleven minutes for you to approve something.

Which does Minnie work with?

All of them, and that was the point.

Minnie is a free macOS desktop pet that drives the agent CLI you already have. She does not ship a model and does not resell one — the agent keeps its own login and does all the reasoning, so there is no second AI bill regardless of which you use. The reasoning behind that is in Why Minnie is free: bring your own AI agent.

Where the agents differ is the consent flow. Minnie’s out-loud “can I do this?” is built on Claude Code’s permission protocol, so today that experience is Claude Code only. Gemini CLI, Codex, Antigravity and custom command-line agents run under Minnie too — you still get the voice input, the visible state and the face — they just run more autonomously, because there is no equivalent prompt for her to speak.

There is a longer version of that, including what “custom CLI” means in practice, in Claude Code, Gemini CLI or Codex: one desktop pet for all.

If you can only pick one

• You mostly write code and want the agent to do the work: Claude Code.
• You mostly review code, or you live in CI: Codex.
• You are reading your way into something huge: Gemini CLI.
• You already pay for one of the three parent subscriptions: start with that one. The gap between these tools is smaller than the gap between using one and not.

New to all of this? Install Claude Code on a Mac, start to finish is the ten-minute version.