Gemini CLI

Why it's ranked here

Google backs this one directly, and it shows in the plumbing: a real monorepo with many separate test suites, install paths for every major package manager, and a permissive Apache license that puts no strings on redistribution or forking. The extensibility story is concrete too, not just a slogan: a documented plug-in protocol and a shipped GitHub Actions integration for code review and issue triage. What holds it back from a higher slot is that the terminal-agent category itself is crowded and undifferentiated at the interface level; this one wins on backing and breadth of tooling, not on a distinctive interaction model.

What's good

The scripting story is well thought out: a one-shot mode returns plain text, but you can also ask for structured JSON or a live stream of events, which matters if you're wiring this into a CI pipeline rather than typing at it interactively. Startup is deliberately kept light: the process that first launches avoids importing the heavy dependencies until it actually needs them, then hands off to a full worker. The test setup goes well beyond unit tests, with separate suites for memory behavior, performance regressions, and multiple sandbox backends, which suggests the maintainers care about the agent not degrading quietly over time.

Tradeoffs

Letting an agent run shell commands and edit files on your machine is inherently risky, and sandboxing here is opt-in rather than automatic, so a careless setup can let the model touch more than intended. The install footprint is heavier than a typical CLI too, with a stack of platform-specific native add-ons for terminal handling that can be a source of build friction on less common systems. There's also a nightly release channel that ships changes from the main line of development before they've been fully checked, which is fine for testing but not something to point at production workflows.

How to use it well

This fits developers who already work from a terminal and want an assistant that can read a whole codebase, make edits, and run commands rather than just chat. It also fits well in CI and scripting workflows, since the structured output modes let you parse its answers programmatically instead of scraping text. Stick to the stable release channel for anything you depend on, turn on sandboxing before letting it run shell commands, and use the plug-in protocol to connect it to your own internal tools rather than writing one-off wrappers. It is not the choice if you want a model-agnostic assistant: it only talks to Gemini models.

Observed

License

Apache 2.0 license, per the README's license badge and text.

Distribution

Distributed as an npm package (@google/gemini-cli), installable via npm, Homebrew, MacPorts, or Anaconda.

Runtime requirement

Requires Node.js 20 or newer, per the engines field in package.json.

Packaging

Packaged as pure ESM (type: module in package.json).

Interface

Primary interface is a CLI binary; also supports non-interactive invocation with structured JSON or streaming-JSON output for scripting.

Extensibility

Extensible through the Model Context Protocol (MCP) for connecting external tool servers.

Platform support

Optional native dependencies ship separate pseudo-terminal binaries for macOS (arm64 and x64), Linux (x64), and Windows (arm64 and x64), indicating cross-platform desktop support.

Repository structure

Organized as an npm-workspaces monorepo (packages/*).

Test suite

Repository includes dedicated integration, memory, and performance test suites beyond unit tests.

Contribution process

Contribution workflow requires a signed Google CLA and includes an in-repo automated PR-review helper script.