Google previews Gemini API CLI (gemini-api), a zero-dependency Go binary that lets Claude Code, Codex, Cursor and other coding agents generate media, transcribe and run managed agents from the shell

Background and pain points

Coding agents increasingly need to do more than write code: make a hero image and voiceover for a landing page, read a PDF, turn a meeting recording into captions, or hand research to a background agent. Until now an agent calling Gemini usually guessed model IDs, wrote a throwaway SDK script or a long curl command, and often dumped base64 images into its context. Google now describes a three-layer stack: the Gemini API itself, the Gen AI SDKs (Python, TypeScript, Go, Java) for production backends, and this new gemini-api CLI for terminals and agents.

How it works

gemini-api is a stateless single Go binary; each invocation runs on its own while the host agent supplies context, picks the model and decides the next step. Commands fall into Create (generate, image, tts, music, video), Understand (analyze, transcribe), Manage (agent, files, models and more) and Advanced (environments, credentials, webhooks, triggers). The whole interface is discoverable from the binary: --usage prints the full command tree as KDL, --schema prints the JSON Schema of request bodies, and --dry-run validates flags and previews the HTTP request with credentials redacted, without network access or the OS keychain. Media commands write files to disk and print only the path; managed-agent runs return an interaction ID to collect results later.

Benchmarks

This is a developer tool release and Google has published no performance or accuracy benchmarks. Verifiable facts: the GitHub repo google-gemini/gemini-api-cli is Apache 2.0, the latest release v0.4.3 was published on 2026-10-19, and the official docs label it experimental and not a supported Google product. Output quality depends on the default models underneath (such as Gemini 3.8 Flash and Nano Banana 2.1); the CLI does not change model quality. It is also a different product from the existing Gemini CLI interactive coding agent, and some developers have already found the naming confusing.

Getting started

On macOS or Linux install with the official install.sh, or run go install github.com/google-gemini/gemini-api-cli/cmd/gemini-api@latest, then create an API key in AI Studio and export GEMINI_API_KEY. Check the setup offline with gemini-api generate "Hello" --dry-run, then drop --dry-run for a real call. Google suggests adding a short block to AGENTS.md, CLAUDE.md or .cursorrules telling the agent to read --help or --usage, preview with --dry-run, and keep the returned file paths and interaction IDs. For production backends the Gen AI SDKs remain the recommended path; the CLI fits shell pipelines, CI and coding agents.