Google releases Gemini API CLI: a command-line tool with no memory, designed for Agents

Google released the official Gemini API CLI. Philipp Schmid announced the news on X, and the key point is clear: it is for use by existing Agents, such as Claude Code, Codex, Cursor, Antigravity, OpenClaw, Hermes. You can also use it yourself. It is not a new coding Agent.

Functionally, this CLI covers common multimodal needs:

Generate images, video, audio, and music.
Analyze PDFs, images, audio, video, and YouTube links.
Transcribe recordings into md / json / srt formats.
Run Deep Research in the background.

The really interesting part is its design: stateless. No memory, no chat mode. The Agent is responsible for context and planning, while gemini-api is only responsible for sending API calls. Context stays on the Agent side, and the CLI stays thin and clean. This is a different approach from AI chat tools aimed at general users. The CLI is designed for Agents; it is self-describing, has --help, and has --dry-run. Agents can see its capabilities through help and check calls through dry-run.

Background: Google AI Studio is Google's model development platform, and the Gemini API is called from there. An Agent can be understood as an AI assistant, such as the AI coding assistants in tools like Claude Code and Cursor. Now this CLI connects the API to the command line, so developers can hand it directly to these tools and let them execute calls for you.

Demo video:

You can get started on the official Google AI Studio page: https://aistudio.google.com/learn/introducing-the-gemini-api-cli?e=0

Or try it directly:

gemini-api --help
gemini-api --dry-run

--help outputs command descriptions, and --dry-run previews a call. For readers unfamiliar with the command line, it is just a tool installed in the computer terminal; enter commands to call the Gemini API. Because it is stateless, each call must include the question or file; context is managed by the Agent, so this CLI itself is very light.

This design approach is closer to Agent-first: the model does not retain context, and context is managed by the Agent. For developers, this lowers the difficulty of integrating Gemini into Agents. For ordinary users, it is a more basic interface, and basic means flexible.