Google Previews Gemini API CLI

Google released an experimental preview of Gemini API CLI on the 23rd local time, letting users run Gemini API features through terminal commands. The tool is available as a binary that requires no separate dependency installation and runs on macOS, Linux and Windows. Users can also install it from source. The current version is 0.4.x and is available on GitHub. Google said it will develop the tool based on developer feedback.

The CLI is designed to reduce the need for AI agents to look up model identifiers or repeatedly use one-off SDK scripts and complex curl commands. Google also cited the difficulty of handling generated media as long strings. The company said typed client libraries suit production services, while developers working in terminals and agents that run commands autonomously need a stateless interface with easily understood inputs and commands. The CLI simplifies model calls and result handling through predictable input and output formats.

Google’s developer tools have three layers. The Gemini API handles inference, multimodal analysis, media generation and agent execution. The Google Generative AI SDK supports Python, TypeScript, Go and Java for production services and backend development. The CLI gives developers and shell-based agents direct terminal access to the API. Commands can be used to explore models, test prompts, generate images, videos and music, analyze multimodal content and transcribe speech.

Media generation is available through the Nano Banana 2.1 image model, Gemini Omni 1.1 Flash video model and Lyria 3.5 music model. Gemini 3.8 Flash TTS supports speech synthesis. It can generate conversations with multiple speakers, while Lyria can create instrumental tracks for background music. Image and video results are saved as files, and their paths appear in the output so they can be used in later steps without conversion.

The CLI can summarize and analyze PDF, image and audio files, transcribe speech and analyze YouTube videos. Transcriptions can be saved as SRT subtitle files with timestamps. Agents that can run shell commands can use the tool outside Google’s own development environment; Google cited Claude Code, Codex, OpenClaw, Hermes and Cursor as examples. It is also usable in Antigravity, where agents can call the CLI for media generation or document analysis and pass a result file or task identifier to the next step. Native integration with Antigravity is planned for later.

Managed agents that run in the background can handle multistep research or coding tasks. Users can delegate work to agents such as Deep Research, check progress with an interaction identifier and receive structured results. The CLI can show commands and options offline, display request formats with “--schema” and validate requests without a network connection using “--dry-run.” It also supports JSON and TOON output, plus data extraction with “--jq.” After installation, an API key can be registered through an environment variable, the operating system’s keychain or a configuration file; interactive setup is also available. The CLI is released under the Apache 2.0 open-source license.

Google released an experimental preview of Gemini API CLI on the 23rd local time, letting users run Gemini API features through terminal commands. The tool is available as a binary that requires no separate dependency installation and runs on macOS, Linux and Windows. Users can also install it from source. The current version is 0.4.x and is available on GitHub. Google said it will develop the tool based on developer feedback.

The CLI is designed to reduce the need for AI agents to look up model identifiers or repeatedly use one-off SDK scripts and complex curl commands. Google also cited the difficulty of handling generated media as long strings. The company said typed client libraries suit production services, while developers working in terminals and agents that run commands autonomously need a stateless interface with easily understood inputs and commands. The CLI simplifies model calls and result handling through predictable input and output formats.

Google’s developer tools have three layers. The Gemini API handles inference, multimodal analysis, media generation and agent execution. The Google Generative AI SDK supports Python, TypeScript, Go and Java for production services and backend development. The CLI gives developers and shell-based agents direct terminal access to the API. Commands can be used to explore models, test prompts, generate images, videos and music, analyze multimodal content and transcribe speech.

Media generation is available through the Nano Banana 2.1 image model, Gemini Omni 1.1 Flash video model and Lyria 3.5 music model. Gemini 3.8 Flash TTS supports speech synthesis. It can generate conversations with multiple speakers, while Lyria can create instrumental tracks for background music. Image and video results are saved as files, and their paths appear in the output so they can be used in later steps without conversion.

The CLI can summarize and analyze PDF, image and audio files, transcribe speech and analyze YouTube videos. Transcriptions can be saved as SRT subtitle files with timestamps. Agents that can run shell commands can use the tool outside Google’s own development environment; Google cited Claude Code, Codex, OpenClaw, Hermes and Cursor as examples. It is also usable in Antigravity, where agents can call the CLI for media generation or document analysis and pass a result file or task identifier to the next step. Native integration with Antigravity is planned for later.

Managed agents that run in the background can handle multistep research or coding tasks. Users can delegate work to agents such as Deep Research, check progress with an interaction identifier and receive structured results. The CLI can show commands and options offline, display request formats with “--schema” and validate requests without a network connection using “--dry-run.” It also supports JSON and TOON output, plus data extraction with “--jq.” After installation, an API key can be registered through an environment variable, the operating system’s keychain or a configuration file; interactive setup is also available. The CLI is released under the Apache 2.0 open-source license.