Gemini CLI
A free, open-source terminal agent that reads your project's files and runs coding tasks using Google's Gemini models.
What you'd use it for
Getting a coding agent working in your terminal on a plain Google login, without an API key, a card on file or a subscription.
- Reads and edits files across a project to make multi-file code changes from a plain-language request
- Runs shell commands as part of a task and feeds the output back into its own reasoning
- Searches the web mid-session to pull in current documentation or examples
- Holds up to a 1M-token context window, so a large codebase can stay in view at once
- Being open source, lets you read exactly what a tool call is about to do before it runs
The verdict
Gemini CLI is the cheapest honest way to find out whether a terminal coding agent suits how you work. Signing in with an ordinary Google account gives you 1,000 model requests a day at no cost, which is real daily development rather than a taster, and the CLI is open source so you can read what a tool call will do before it runs. The catch is the shape of the ceiling: requests, not tokens, so one sprawling refactor and one one-line fix cost the same, and a request that returns something useless still counts. Past 1,000 you are choosing between a Google AI subscription, pay-as-you-go API billing, or a Cloud Code Assist seat, and at that point the comparison against a paid competitor is on the agent's merits with the price advantage gone.
What it costs
| Tier | Price | For | Includes |
|---|---|---|---|
| Free -- sign in with Google | $0 | Anyone trying the tool, and most solo daily use. | 1,000 model requests per user per day, across whichever Gemini models the CLI picks. No API key, no card. |
| Free -- Gemini API key | $0 | Developers who would rather authenticate with a key than a Google login. | 250 model requests per user per day, Flash model only. A quarter of the Google-login allowance, on a weaker model. |
| Google AI Pro / Ultra | Subscription, priced by Google separately from the CLI | A developer who keeps hitting 1,000 requests a day. | 1,500 requests per day on Pro, 2,000 on Ultra, using the same Google sign-in. |
| Pay-as-you-go (API key) | Usage-based, standard Gemini API rates | Usage beyond the free daily caps with no fixed monthly commitment. | Billed per request once the free allowance is used. |
| Code Assist Standard / Enterprise | Contact Google Cloud sales | Organisations wanting centrally billed, licensed seats. | 1,500 requests per user per day on Standard, 2,000 on Enterprise, with Workspace and Cloud administration. |
Read from Gemini CLI's own pricing page on 25 Sep 2026. Vendors change prices without notice.
What you need to start
- Node.js 20 or newer, plus npm
- A Google account to sign in with -- no API key, card or subscription for the free allowance
- A web browser on the same machine as the terminal, since sign-in completes in the browser
- A Google Cloud project only if you are signing in with a company, school or Workspace account, or using a Code Assist licence
- A terminal you are comfortable working in -- there is no graphical interface
Minutes. Install with npm, run it in a project directory, sign in with Google in the browser window it opens, and it is reading your files.
Your first run Install with npm, run gemini in a project directory, and pick "Sign in with Google" — a few minutes later you have a session that will read and edit a real file in that directory, with no API key and no card.
Where it breaks
The caps count requests, not tokens, so a trivial question and a whole-repository refactor cost the same one request -- and a request that comes back useless is still spent. Google publishes a per-minute limit but not its number, saying only that requests are limited per user per minute and subject to availability at busy times, so short bursts can be throttled without a figure you can plan against. The daily allowance is also per Google account rather than per project, so a team sharing one login reaches 1,000 far sooner than the number suggests.
Installs via npm as a command-line tool; runs in your terminal and reads/edits files in whatever project directory you point it at, the same model as [Claude Code](/tools/claude-code/) or [Codex](/tools/openai-codex/).
First entry in this reference. There is no previous check to compare it against.
When not to use it
Not for a team that wants one bill and one admin console for everyone's usage -- the free allowance is per Google account, so it works by each person signing in as themselves, and central billing means moving to Code Assist licences or Vertex AI. Also not for anyone who will not work in a terminal: there is no graphical interface, and no plan adds one.
Gemini CLI is Google’s answer to the same job Claude Code and Codex do: a terminal agent that reads your actual codebase, makes changes across files, and runs commands, rather than a chat window you paste code snippets into. The distinguishing fact for a small business is the price: signing in with an ordinary Google account gives you 1,000 model requests a day at no cost, which is enough for real daily development work before you would need to pay anything.
Open source, not just free
Unlike the closed harnesses it competes with, the CLI itself is open source, published on GitHub under Google’s name. That matters less for day-to-day use than it does for trust and longevity — you can read exactly what it does, and it isn’t going to disappear if a single vendor’s pricing changes.
The two free paths are not equal, and the obvious one is wrong
There are two ways in at no cost, and the one a developer reaches for first is the worse of the two. Signing in with a Google account gives 1,000 model requests a day across whichever Gemini models the CLI selects. Authenticating with a Gemini API key from AI Studio instead gives 250 a day, Flash only: a quarter of the allowance on a weaker model. The API key feels like the more technical, more capable route and is the one to avoid unless something in your setup rules out an interactive browser login.
Past either ceiling the options are a Google AI subscription (1,500 requests a day on Pro, 2,000 on Ultra, same sign-in), pay-as-you-go API billing, or a Code Assist seat through Google Cloud for a team that wants one bill.
Requests, not tokens, which changes how you use it
The unit to watch is the request. A one-line question and a refactor spanning twenty files each cost exactly one, and a request that comes back wrong is still spent. That inverts the habit most people bring from token-metered tools: batching a real piece of work into a single well-framed prompt is cheaper here than a conversation of small corrections, and iterating your way to an answer is the expensive path. Google also limits requests per minute without publishing the number, so a burst can be throttled with nothing to plan against.
Where it fits
For a solo founder or small team, this is a genuinely free way to find out whether a terminal coding agent suits you at all, before paying for Claude Code or Cursor. Try it first if budget rather than capability is the current constraint. The honest comparison comes later: once you are past 1,000 requests a day and paying either way, the price advantage is gone and the choice is on the agent’s merits.
Alternatives
Websites & appsHarnessClaude Code
A terminal-based agent that reads your whole project and makes changes across files on instruction, sold as part of a Claude subscription.
Websites & appsHarnessCursor
A code editor where you describe the change you want in plain English and it edits the files directly, used by non-developers for small internal tools.
Websites & appsHarnessOpenAI Codex
A cloud coding agent that takes a task, works in its own sandbox without supervision, and comes back with a finished pull request for you to review.
Get Gemini CLI
Buying the tool is not the hard part.
The ebook is about what to hand over, in what order, and how you find out when it goes wrong. One payment, no subscription.