What is claude-antigravity-agents?

This project is a skill for Claude Code, meaning a single instruction file that teaches Claude Code a new capability rather than a standalone application. Once installed, it lets Claude Code hand off heavy jobs, like a full repository audit, a large refactor, or a research sweep, to Google's Antigravity CLI, a separate terminal coding agent that runs on Gemini and other Google models. The idea is that instead of Claude Code doing every part of a task inside your current session, burning your context and subscription tokens along the way, it can write a self contained brief describing the job, launch it as a background Antigravity process, and keep working on whatever you actually asked it to do. When the background job finishes, Claude Code reads the result and checks it, running things like a git diff, a typecheck, or a build, before treating any of the returned work as trustworthy. This also lets you get a genuinely independent second opinion, since Antigravity jobs run on a different model than Claude itself, which can be useful for reviewing a design decision or catching bugs a single model might miss. The project sets a firm rule for safety: an Antigravity job is never allowed to write to the same files Claude Code is actively editing at that moment. Read-only jobs like code review or security analysis run in a sandboxed mode alongside Claude, while jobs that write code run in an isolated git worktree or a separate repository instead. Using it requires installing the Antigravity CLI and completing a one time Google sign-in, then copying this project's skill folder into your local Claude Code skills directory. After that, you simply ask Claude Code in plain language to delegate something, and it decides the job type and handles the rest. It is released under the MIT license.