Google's gemini-api CLI hands Gemini media and managed agents to Claude Code, Codex and other shell agents

Google has released an experimental preview of the Gemini API CLI, a command named gemini-api. The launch post, by Dmitry Lyalin and Philipp Schmid of the AI Studio team, is dated October 19 and calls it “a zero-dependency binary” that brings Google’s reasoning models, media generation, multimodal analysis and Managed Agents into the terminal and into coding agents. The repository is Apache 2.0, and the latest release is v0.4.3.

Do not confuse it with Gemini CLI, Google’s interactive coding agent. This one is a stateless tool meant to be called by an agent you already run.

What it does

The README’s opening line is “Gemini models and managed agents for the agent you already use.” It names Codex, Claude Code, Antigravity, OpenClaw and Hermes, and the launch post adds Cursor. You set GEMINI_API_KEY, put the binary on PATH, and the agent calls commands such as:

• generate and analyze for text and file questions, defaulting to Gemini 3.8 Flash
• image for generation and editing, defaulting to Nano Banana 2.1
• video, music and tts, defaulting to Gemini Omni 1.1 Flash, Lyria 3.5 and Gemini 3.8 Flash TTS
• transcribe, which can write SRT subtitles
• agent run, which starts a background Managed Agent such as Deep Research and returns an interaction ID to poll later

Media commands write files straight to disk and print only the file path, so base64 payloads stay out of the agent’s context window.

Built for agents to read

Commands describe themselves offline. --help and --usage print the command tree (the latter in machine-readable KDL), --schema prints request schemas, and --dry-run shows the resolved HTTP request with credentials redacted, without a network call. Output can be JSON or compact TOON, with inline --jq filtering. The README says exit codes separate success (0), runtime or API failures (1), usage errors (2) and auth failures (3), and --agent-mode turns errors into structured JSON.

Keys can live in the environment, the OS keychain or ~/.config/gemini-api/config.yaml. Install is a curl script for macOS and Linux, a PowerShell script for Windows, or go install.

Caveats

The README says the tool is in beta and evolving alongside the Gemini Interactions API, that breaking changes may occur, and that CI users should pin a release. Calls bill against your Gemini API key like any other API use. Google says native integration with Antigravity is “coming soon”; you can already call the CLI from Antigravity’s shell.

Why it matters

Coding agents increasingly reach for outside models for images, voice or research. Google’s answer is a plain shell command that any agent with a terminal can already run, with no SDK code to write.