What Gemini CLI does on a Mac and how it differs from other coding agents

Anyone searching for Gemini CLI in late 2026 is looking at a tool whose ground shifted underneath it. The project is still on GitHub, the install command still works, and the documentation site is still live. What changed is who is allowed to send requests through it. Before deciding whether to install it on a Mac, it pays to know which side of that line your Google account sits on.

The change that happened on June 18, 2026

Google announced on May 19, 2026 that Gemini CLI was being folded into Antigravity, and that a new terminal client called Antigravity CLI would take its place. The announcement was specific about dates and about who it affects.

From that post: on June 18, 2026, Gemini CLI and the Gemini Code Assist IDE extensions stopped serving requests for Google AI Pro and Ultra subscribers, and for anyone using Gemini Code Assist for individuals free of charge. The same cutoff applied to Gemini Code Assist for GitHub, with no new installations on GitHub organizations after that date.

Organizations were carved out. If a company holds Gemini Code Assist Standard or Enterprise licenses, or uses Gemini Code Assist for GitHub through Google Cloud, access continues and Google says it will keep supporting the tool with current models. Gemini CLI also remains reachable with paid Gemini and Gemini Enterprise Agent Platform API keys.

The documentation site carries the same message in a banner at the top of every page: unpaid tier and Google One users were moved to Antigravity CLI on June 18, 2026.

This matters because the repository README still describes the old consumer path, including a free tier of 60 requests per minute and 1,000 requests per day through a personal Google account. That text predates the transition. Reading only the README leads to a plan that no longer matches what the service will accept.

So the practical question on a Mac is not how to install Gemini CLI. It is which credential gets pointed at it. A paid API key or an organization license keeps the tool working. A personal Google account on the consumer tiers does not, and that path leads to Antigravity CLI instead.

Installing and authenticating on macOS

Installation itself is unremarkable, which is a compliment. Gemini CLI is a Node package published as @google/gemini-cli, and the project lists four routes.

npx @google/gemini-cli          # run without installing
npm install -g @google/gemini-cli
brew install gemini-cli
sudo port install gemini-cli

Release channels are published on a fixed schedule. Stable ships weekly on Tuesdays at 20:00 UTC and is the promotion of the previous week's preview plus fixes. Preview ships Tuesdays at 23:59 UTC. Nightly ships daily at 00:00 UTC from whatever is on the main branch, with validation still pending. The tags map directly to install commands: @latest, @preview, @nightly.

Authentication is where the decision from the previous section lands. The project documents three routes:

Route	Who it fits now	Google Cloud project needed
Sign in with Google	Organization accounts with a Code Assist Standard or Enterprise license	Yes for company, school, or Workspace accounts
Gemini API key	Developers who want model control and pay for usage	No
Vertex AI	Teams already running on Google Cloud	Yes

For an individual on a Mac who is not attached to an organization license, the API key route is the one that still leads somewhere. The key goes in an environment variable, which means it can live in a shell profile or a project-level environment file rather than being typed each session.

export GEMINI_API_KEY="your-key"
gemini

The tool is Apache 2.0 licensed and the source is public, which is worth weighing separately from the hosted service. The client keeps working as a piece of software; what the transition changed is which backends will answer it.

What the agent actually does inside a folder

Gemini CLI is a terminal agent in the same family as the other tools in this category. It reads and edits files, runs shell commands, fetches web pages, and can ground answers in Google Search results. It speaks MCP, so custom tool servers plug into it the same way they plug into other agents.

Three behaviors are worth knowing before the first session.

It takes the current directory as its world. Running gemini in a project folder gives the agent that folder as context. Extra directories can be added at launch with --include-directories ../lib,../docs, which is how a monorepo or a docs folder living outside the project gets pulled in.

Project instructions live in GEMINI.md. A context file in the repository tells the agent the conventions of that project, in the same spirit as AGENTS.md for Codex or CLAUDE.md for Claude Code. Every agent in this category has converged on the same idea: the folder carries its own instructions, so the operator stops repeating them.

It runs non-interactively. A single prompt can be passed with -p, and the output format switched to JSON for parsing or to newline-delimited streaming events for monitoring a long run.

gemini -p "Explain the architecture of this codebase" --output-format json

Conversation checkpointing lets a long session be saved and resumed, which matters more than it sounds when a refactor spans two working days.

Where search grounding changes the answer

One feature separates Gemini CLI from most of its neighbors: it can ground a response in Google Search results rather than answering from training data alone. That sounds like a marketing line until it hits a specific class of task.

Terminal agents fail in a predictable way when a library has changed. The model produces code for the version it learned, the project is on a newer one, and the error that comes back is a missing method rather than an obvious mistake. An agent that can search before answering has a route out of that trap, provided the operator asks it to check.

The same mechanism helps with anything dated: a deprecation notice, a provider that renamed a plan, an installation step that moved. Worth stating plainly, though, is that grounding is not verification. A search result can be wrong, outdated, or written by someone guessing. Treat retrieved text the way a careful reader treats any web page, especially when the agent is about to act on it.

Other agents have reached the same conclusion by different routes. Codex CLI serves web results from a maintained cache by default and switches to live results only under a full-access setting, specifically to reduce exposure to instructions hidden in arbitrary pages. Reading the documentation of whichever agent is in use, on that one point, is time well spent before granting it network access.

How it compares with the other terminal agents

Three agents dominate this space on a Mac, and they differ less in capability than in how they are sold and where the work is allowed to run.

	Gemini CLI	Claude Code	Codex CLI
Install	npm, Homebrew, MacPorts, npx	Native installer, Homebrew, npm	Standalone script, npm, Homebrew
Source	Open, Apache 2.0	Closed	Open, Apache 2.0
Project instructions	GEMINI.md	CLAUDE.md	AGENTS.md
Consumer plan access	Ended June 18, 2026	Pro, Max, Team, Enterprise	Free, Go, Plus, Pro
Entry price	Paid API key or org license	20 USD per month for Pro	0 USD on Free, 8 USD on Go
Sandboxing	Configurable	Configurable	OS-enforced, network off by default

The pricing column deserves a caveat. Anthropic lists Claude Code as included in Pro at 20 USD per month billed monthly, or 17 USD per month when billed annually at 200 USD up front, with Max starting at 100 USD. OpenAI lists Codex on the Free plan for quick tasks, Go at 8 USD, Plus at 20 USD, and Pro from 100 USD. All three change these numbers regularly, and taxes are not included.

A second caveat applies to the install column. Being a Node package makes Gemini CLI easy to place on any machine that already has a runtime, and it also ties the tool to that runtime. Claude Code and Codex CLI both ship standalone binaries that do not call Node at all, so a Node upgrade cannot break them. On a Mac where several projects pin different Node versions through a version manager, that difference shows up as an occasional broken global command rather than as anything dramatic.

The more useful way to read the table is by risk. An agent that is open source and authenticates with a key you control can be pointed at a different backend later. An agent bundled with a consumer subscription is cheaper to start and harder to move. The Gemini CLI transition is a live demonstration of why that distinction is not academic.

The constraint none of these tools solve

Every terminal agent in this category assumes the same thing: that the work already lives in one folder, and that the operator is sitting in a terminal opened at that folder.

Real work rarely looks like that. A single piece of work spreads across a repository, a folder of reference PDFs, an export sitting in Downloads, and media on an external drive. The agent sees one of those places. Getting the rest in front of it means opening Finder, finding the path, switching to the terminal, typing cd, and starting again. That round trip repeats dozens of times a day and none of it is visible in a benchmark.

Two changes reduce it. The first costs nothing: pick a folder per piece of work, keep the agent started there, and write the project's conventions into the context file so they stop being retyped. The second is structural, and it is to stop running the file browser and the terminal as separate applications. A file manager with a built-in terminal opens a shell in whatever folder is selected, so the agent starts in the right place without a cd, and the files it rewrites stay visible in the same window. The comparison with other file managers sets out what that changes in practice.

The other half of the problem is time. An agent run takes minutes, then stops to ask a question, and the answer waits until someone sits back down. Being able to read a running terminal from an iPhone or iPad turns that dead time into a short reply from wherever you are.

What to change first

If the Google account in question is a personal one, the honest first step is to check which client the credential still works with before investing a week of habits in it. If the account is covered by an organization license, Gemini CLI remains a reasonable terminal agent, and the open license means the client is not a dead end.

Either way, the change with the longest payoff is the one underneath the agent: one folder per project, its conventions written down in the context file, and a window where the folder and the shell sit together instead of across an application switch. Atriens was built around that last point, and everything it does on the Mac itself is free to try.