AI for code tower · floor

Gemini CLI: open-source client, three sets of terms

Google describes it in six words on the front page of the repository: “an open-source AI agent that brings the power of Gemini directly into your terminal”. Both halves of that are true, and they are true of different things.

That distinction is not a criticism. Google says it plainly on the same page, and it follows from what the program is: a client. The model is somewhere else, and the client is the part that was opened. But it changes what the word “open” settles. It settles what the program does with your files and how it behaves, because you can read it. It does not settle what happens to what you send. For that you have to know which door you came in through.

That table is the most useful thing on this page for anybody deciding whether to point this at work code. The same command, typed the same way, sits under Google Terms of Service, Gemini API terms, or Google Cloud Platform terms depending on which sign-in was used. That choice is made once, during setup, and it is not restated afterwards — so if the answer matters, the thing to go and check is how you authenticated.

The other number people arrive looking for is the limit, and Google publishes it as a table rather than a headline.

Two details in that table are worth pulling out, because the numbers alone do not carry them. The free route through a Google account gives four times the daily requests of the free route through an API key — and the difference is not only volume. Google says the API-key free tier makes “model requests to Flash model only”, while the Google-account route spreads requests “across the Gemini model family as determined by Gemini CLI”. So one free tier is larger and picks its own model; the other is smaller and fixed.

And a caveat attached to all of it: requests “are limited per user per minute and are subject to the availability of the service in times of high demand”. A published ceiling is not a guarantee of a floor.

Getting it, in Google’s own words. Three routes are listed: run it without installing with npx @google/gemini-cli, install it globally with npm install -g @google/gemini-cli, or on macOS and Linux with brew install gemini-cli. The first of those is the one to know about — it lets you try the thing without leaving anything behind on the machine, which for a tool that reads your files and runs shell commands is a reasonable way to start.

What the tool can reach once it is running is stated plainly in the same README, and it is a wider set than “a chat window in a terminal” suggests: Google Search grounding, file operations, shell commands, web fetching, plus support for MCP — the Model Context Protocol — “for custom integrations”. Two items on that list deserve to be read slowly. File operations and shell commands mean the agent is not only answering; it is acting, on the machine you are sitting at.

That is worth being clear-eyed about, and it is stated by Google rather than inferred here: file operations and shell commands are on the published list of what the tool can do. What no documentation settles is how much room to give it. This page’s own suggestion, offered as such: point it first at something you could throw away and rebuild, rather than at the repository you could not.

One last thing Google publishes that is easy to miss, and it is a control rather than a warning. “Your Gemini CLI Usage Statistics are handled in accordance with Google’s Privacy Policy” — and there is an opt-out: the documentation links to a Usage Statistics configuration page for turning the sending off. That is separate from what happens to your prompts, which is governed by whichever of the three sets of terms applies to you.

What this page will not tell you is whether Google trains on what you send. That answer lives in the individual privacy notices for Gemini Code Assist, the Gemini API and Vertex AI respectively — three different documents, for three different sign-in routes, and it is those that have to be read rather than a summary of them.

The floors below take it three ways: the licence and the terms, the limits, and the wider question of what it means to let a program act on your machine.

What this tower will not do

It will not tell you the open licence covers your data. Google states in the same document that the terms of the services being called apply to that, and names three different sets of them.

It will not quote one free limit. Google publishes a table with different numbers for different sign-in routes, and picking one figure out of it would be right for some readers and wrong for the rest.

And it will not judge whether the tool is any good. That needs running it against real work for long enough to have an opinion worth publishing, which this site has not done. What holds instead is simple: the description of Gemini CLI as an open-source AI agent, the Apache 2.0 licence and the statement that the terms of the services being accessed apply separately, the three authentication routes and their respective terms, the note about organisation-managed accounts, the daily request table, the difference between the two free tiers including the Flash-only limitation, the high-demand caveat, the prohibition on third-party access to the underlying services, the installation commands, the built-in tool list and the usage-statistics opt-out are quoted from Google’s own Gemini CLI documentation, listed below.