Google’s new gemini-api CLI is built for agents and CI, not for chatting

Google has released the Gemini API CLI (gemini-api) in beta, a stateless, zero-dependency command-line client for coding agents, scripts and CI pipelines. The entry appeared in the Gemini API release notes dated October 27, 2026.

According to the changelog, the tool works on the Interactions API. It can generate text and media, analyze files and run Gemini managed agents. The release notes point to a Gemini API CLI guide for setup.

What the tool does and how to install it

Google’s CLI guide groups the commands by task. The creation commands are image, tts, music, video and generate. For understanding content there are analyze and transcribe. A separate set handles managed agents: agent (with run, status and cancel), plus environments, credentials, webhooks and triggers. Housekeeping commands include configure, whoami, files and models.

It runs on macOS, Linux and Windows. Installation is by a shell script on macOS and Linux, a PowerShell script on Windows, go install, or prebuilt archives from GitHub Releases, which the guide recommends for pinning versions in Docker and CI. Authentication uses the GEMINI_API_KEY environment variable. Locally, gemini-api configure --api-key stores the key in the OS keychain, or in ~/.config/gemini-api/config.yaml with --no-keyring.

The guide calls it an “open-source beta release” but does not name a license. It gives no separate pricing for the CLI and refers readers to the general Gemini API pricing page, so usage is presumably billed like any other API call.

Stateless by design, with flags for machines

The defining choice is statelessness. Per the guide, commands keep no hidden local state. Multi-step work is chained explicitly with --previous-interaction-id, an uploaded file URI, or a background interaction ID. Media commands write files to disk and print only the path, which keeps output short for an agent reading a terminal.

Several flags target automation: --schema and --usage let an agent inspect a command, --dry-run previews a call, and -o json, -o toon and --jq shape the output. Commands never prompt unless --interactive is set. Exit codes are fixed: 0 for success, 1 for a runtime or API error, 2 for a usage error and 3 for an authentication failure.

Limits are spelled out too. Inline files are capped at 14 MB raw (20 MB with base64); larger ones need files upload, and uploaded files are deleted after 48 hours. There are no bidirectional WebSockets, so real-time streaming still needs the Live API and the GenAI SDKs.

A different tool from Gemini CLI

The name invites confusion with Google’s existing Gemini CLI, which has kept shipping releases this week, including v0.66.0 on October 27. Google’s guide separates them: “gemini-api is a stateless API client, not an interactive coding agent like @google/gemini-cli.” In other words, gemini-api is meant to be called by an agent, including one built on another vendor’s model, rather than to be the agent.

The release also fits a pattern in the same changelog. The Interactions API stores requests by default and, per its documentation, keeps them for 55 days on the paid tier and one day on the free tier. Setting store=false opts out, but the documentation says it also disables background execution and previous_interaction_id. Anyone chaining CLI calls with --previous-interaction-id therefore depends on server-side storage. Separately, on October 26 Google switched on storage by default for new paid-tier projects using Generate Content, so teams with data-handling rules should check those settings before putting the CLI in a pipeline.

What is still unclear

The changelog entry is short and gives no release date beyond October 27, no version number and no benchmark of how well agents use the tool. The license is not named, and the beta label means commands and flags may change. Claims about suitability for agents come from Google’s own documentation and have not been independently tested. Our earlier coverage of Gemini 4 Argon covers the model that the same changelog lists as generally available since October 13.