Gemini CLI

What it is

Gemini CLI is Google's terminal coding agent, published under Apache-2.0 — the most permissive licence on this board, and a real one: you can fork it, ship it inside something else, and read every line of the loop. It drives Gemini models only, so the openness is in the client rather than the stack.

Its distinguishing feature has always been context. Gemini's very large windows let the agent hold a substantial share of a repository at once rather than rediscovering it file by file, which changes the character of the work: fewer wasted turns re-reading, better answers to questions that span a dozen files, noticeably stronger on the "why is this broken" class of task than on the "write this function" class.

How the loop works

Read, search, edit, run, with shell access and MCP servers for anything beyond the built-ins. Project instructions live in a Markdown file the agent picks up from the repository, and those files nest — a root file for house rules, per-directory files for local ones. There is a checkpointing mechanism so you can roll back a set of edits the agent made, which matters more here than elsewhere because big-context agents tend to make broad multi-file changes rather than cautious single-file ones. It also runs non-interactively, taking a prompt and emitting output, which is what makes it usable in scripts and CI.

Grounding through Google Search is available as a tool. Treat results the way you would treat any other untrusted text entering the context: they are data, not instructions.

Isolation and permissions

The spec says configurable, and that is the honest word. Out of the box you approve tool calls interactively and can allow individual commands for the session. Stronger isolation is available but opt-in: a sandbox mode that runs tool execution inside a container, or OS-level confinement on macOS. Neither is on unless you ask, so the default posture is closer to Claude Code than to Codex CLI.

There is a second, quieter consideration. Large-context agents read a lot of your repository and send it to a model provider on every turn. If parts of the tree must not leave the building, exclusion rules are not an afterthought — configure them before the first run, and verify what is actually being sent.

Who it is for

People working in large or unfamiliar codebases, where understanding is the bottleneck rather than typing; teams already on Google Cloud with Gemini billing in place; and anyone whose licensing requirements rule out proprietary clients. It is a reasonable "second agent" too — worth keeping around for whole-repository questions even if your day-to-day loop is something else.

Limits

• The free tier is gone. It ended 18 Jun 2026. The generous no-cost allowance was most of the reason for its adoption, and its absence makes this a budgeted decision like any other.
• Big context is not free context. Every turn that carries half a repository is billed for half a repository. The failure mode is a surprising invoice, not a surprising diff.
• Gemini models only. Apache-2.0 on the client does not buy you model choice.
• Isolation is opt-in. If a sandbox is a hard requirement rather than a preference, the default configuration is not it.
• Recall degrades before the window ends. A model that accepts a million tokens does not attend to all of them equally; a curated 50k of the right files still beats an undifferentiated dump.

Alternatives on this board

• Codex CLI — open source with the sandbox on by default.
• Amp — the other answer to large codebases, solving it with deep search rather than a large window.
• Claude Code — stronger session and context tooling, proprietary.
• OpenCode — open source and provider-agnostic, so you can point it at Gemini and at everything else.