Gemini CLI: the free-tier workhorse, set up in minutes

Every terminal agent I've onboarded starts the same way: create an account, add a payment method, generate an API key, export it in your shell, and only then do any real work. Gemini CLI's bet is that this friction is the actual adoption problem. You install it with npm, sign in with the Google account you already have, and you're running an agent backed by a serious long-context model. No credit card enters the picture.

Install, sign in, go

npm install -g @google/gemini-cli
gemini

Prefer not to commit to a global install? npx @google/gemini-cli works for a one-off session. Run gemini inside a project directory and it opens an interactive session; first launch walks you through auth. Pick the Google-account option, approve in the browser, done. There are API-key and Vertex paths for teams and CI, but for a first session the personal account is the point: it gets you the free tier.

And the free tier is the story. As of this writing it runs to roughly a thousand model requests a day with a healthy per-minute allowance, backed by Gemini's long-context models and their 1M-token window. Nothing else in the 2026 tool landscape hands out that much agent capacity for a sign-in. If you want the model-family context first, start with the Gemini overview.

The loop and the built-in tools

Strip the branding and Gemini CLI is the standard agent loop: read context, propose an edit or a command, observe the result, repeat. What makes it useful in the first ten minutes is that the toolset ships built in, nothing to configure:

• File tools. Read, write, and edit files, plus glob and grep-style search across the repo.
• Shell. Runs commands (builds, tests, git) with a confirmation prompt before anything executes.
• Web. Fetches pages and grounds answers in Google Search, which earns its keep the first time you debug against a library that shipped breaking changes last month.

Give it something real on day one: ask it to trace a request from route handler to database in a repo you know, or hand it a flaky test and ask why. It reads what it needs, proposes edits as diffs, and asks before touching disk or running anything. Type /tools to see the toolset and /stats to watch your usage; the slash commands are half the learning curve, and there aren't many of them.

The cheapest experiment in agentic coding right now is one npm install and a Google sign-in.

When free is genuinely enough

The free tier fits a real class of work, not just demos: solo development, side projects, spelunking an unfamiliar codebase, a couple dozen agent tasks a day. Interactive use rarely hits the ceiling. Long autonomous runs that burn requests in loops, parallel sessions across several worktrees, and anything wired into CI push past it. There are two honest catches. Under load the free tier can quietly route you to a smaller, faster model than the one you started on. Fine for renames, noticeable on subtle refactors. Read the data-use terms before pointing it at proprietary code, too; free tiers are rarely free in every sense, and a paid API key buys cleaner guarantees.

My advice: if you've been curious about terminal agents but unwilling to fund the experiment, run this one. Give it a bounded, real task on a repo you know well and judge it against your own bar, not the demo reel. If it sticks, you'll eventually feel where the quota pinches, and by then you'll know precisely what paying would buy you.