// the find
google-gemini/gemini-cli
An open-source AI agent that brings the power of Gemini directly into your terminal.
Google's official terminal AI agent built on Gemini, functionally the same category as Claude Code or Aider: a CLI that reads your repo, edits files, runs shell commands, and talks to MCP servers. It's for developers who want a free-tier agentic coding assistant tied to a Google account rather than a paid API key.
The CI/eval setup is unusually serious for a CLI tool — dozens of workflows plus a dedicated evals/ directory with eval files for things like plan mode, subagents, prompt injection via MCP, and memory persistence, which suggests actual regression testing against agent behavior rather than just unit tests. MCP client and server support means tools built for Claude Code's ecosystem are largely portable here too. Headless mode supports structured JSON and newline-delimited streaming JSON output, so it's usable as a real pipeline component, not just an interactive toy. Auth story scales cleanly from a free OAuth login (60 req/min, 1000/day) up through API keys to Vertex AI for enterprise, without forcing a credit card on day one.
The README is straight landing-page copy — emoji bullet lists and feature claims with zero explanation of how the agent loop, tool permissions, or sandboxing actually work; you have to leave the repo for the docs site to understand the architecture. Three release channels (nightly, preview, stable) all cut from main, with the docs explicitly admitting preview 'may contain regressions' — that's a lot of churn to land on 'stable' weekly. Distribution is npm/Docker only, no static binary, so adopting it means pulling in a full Node toolchain and its dependency surface just to run a CLI. Default onboarding path is sign-in-with-Google, which is the easiest route but also the one most likely to send telemetry back to Google by default — worth reading docs/cli/telemetry.md before pointing this at a private codebase.