Google’s new Gemini API CLI is built for AI agents, not for people

Google added the Gemini API CLI (gemini-api) to its Gemini API changelog in an entry dated October 27, 2026. The beta tool is a command-line client for AI coding agents, scripts and CI pipelines, according to the entry.

The entry calls it stateless and free of dependencies. It works with the Interactions API to generate text and media, analyze files and run Gemini managed agents. The changelog links to a CLI guide, the Interactions API overview and a page on managed agents.

One binary with commands for text, media and agents

Google’s CLI documentation lists the commands in groups. The creation set covers generate for text, plus image, tts, music and video. For understanding content there are analyze and transcribe. Agent work goes through agent, which handles running, status checks, cancelling and custom agent management. Further commands cover environments, credentials, webhooks and triggers, along with files, models, configure, auth and whoami.

Installation is by a shell script on macOS and Linux, a PowerShell script on Windows, go install, or prebuilt archives from GitHub Releases. Authentication uses a GEMINI_API_KEY environment variable. On a personal machine, gemini-api configure can store the key in the operating system keychain, or in a config file when --no-keyring is passed. A release version can be pinned with GEMINI_VERSION.

Several features target software callers rather than humans. According to the documentation, media commands write a file to disk and print only its path, --dry-run previews a request without calling the API, and --usage and --schema give an offline reference for commands and requests. Output can be made machine-readable with -o json, --jq and --agent-mode, which also returns structured errors. Exit codes differ for runtime, usage and authentication failures, and --body accepts raw Interactions API JSON for options the flags do not cover.

What the stateless design leaves to the caller

Stateless means every invocation stands alone. The calling agent or script decides the model and supplies any context, and long agent runs are handled through status checks rather than a session held by the tool. For an agent with shell access, that keeps the interface simple: one command in, one result out, with no daemon to manage.

The tool is also narrower than its name might suggest. It is a different product from Google’s interactive Gemini CLI, which has its own GitHub releases; Google’s nightly builds of that tool continued this week. The new client is a thin wrapper over one API.

Limits and open questions for a beta

The documentation sets out several constraints. Inline files are capped at 14 MB raw, or 20 MB once base64-encoded, and larger inputs must go through the Files API. Uploaded files are deleted after 48 hours. The CLI supports HTTP, server-sent event streaming and async polling, but not bidirectional WebSockets, so real-time voice and video still require the Live API. Streaming combined with -o json or --jq requires --stream=false.

Google labels the project an open-source beta in which breaking changes can arrive between minor versions, and it advises pinning a release in automated environments. That matters for the exact audience the tool targets, since a CI job that tracks the latest version could fail after an update. The GitHub repository says the software is under the Apache 2.0 license, and on the day it was read it showed 399 stars and 55 forks. It also says the project is not taking pull requests for now and asks for feedback through GitHub Issues.

The changelog entry and documentation do not say what the CLI costs beyond normal Gemini API usage, which model versions each command defaults to, or how the tool will be supported once it leaves beta. They also give no usage figures, and Google’s description of the tool as suited to agents is its own framing, not something independently tested. The GitHub page showed no published releases when it was read, and a third-party listing cited both v0.4.3 and v0.4.4 as the latest version, so the current version number is unclear.

For developers already building on the Interactions API, the practical change is a ready-made way to call it from a shell, a script or a coding agent without writing a client. Teams that need real-time audio, large files or a stable interface will have to wait or use other routes.