Google's new Gemini API CLI gets one thing right that a lot of agent tooling s… · @willzero · TwKit

@willzero

Google's new Gemini API CLI gets one thing right that a lot of agent tooling still doesn't: generated media goes to disk, and the agent only gets back a file path. Per Google's launch post, gemini-api is an experimental preview, a single Go binary for macOS, Linux and Windows. Any coding agent with a shell can call it. Google lists Claude Code, Codex and Cursor next to its own Antigravity. What one binary covers: • text and reasoning models • image, video, speech and music generation • PDF and media analysis, plus audio transcription to SRT subtitles • background managed agents like Deep Research, which hand back an interaction ID you check later The image, video, TTS and music commands save the file and print only its path. For an agent building a landing page, that means a hero image costs a short line of context instead of a base64 blob. I also like that it explains itself offline. The agent can read the help, usage and schema output, and a dry run prints the resolved request with the key redacted and makes no network call. So it can check a video request before paying for one. Two caveats. It's a v0.4.x preview, and the README warns breaking changes may land between releases and recommends pinning a version in CI. I'd do that. The name is also confusing: several replies are asking how it relates to the older Gemini CLI. It authenticates with a Gemini API key. A few people in the replies are asking whether a Pro or Ultra subscription covers it, and the launch post doesn't address that. If you build tools that agents call, returning a path instead of the payload is the part I'd copy first.

曝光 26 · 评论 0 · 点赞 0 · 书签 0