Antigravity for Claude Code: Claude conducts, Gemini does the heavy lifting
Claude Code plugin: run the Antigravity CLI (Gemini) as a collaborating sub-agent with intelligent model routing across the SDLC. Community project; not affiliated with Google/Anthropic.
At a glance
- What is it?
- A Claude Code plugin that delegates bulk implementation, test generation and research to Google's Antigravity CLI under cost discipline, with measured savings of 27 to 64 percent on a large build.
- Who is it for?
- Antigravity for Claude Code fits teams running large, separable builds through Claude Code who also have Gemini-side access, want cross-model verification and media understanding, and will set their real rates in prices.json before trusting the savings numbers. Skip it on native Windows, for tiny tasks where its own data says delegation loses money, or if adding a second vendor's CLI to your loop is more complexity than the 27 percent is worth.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Shell, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 17, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
Conductor and executor
Antigravity for Claude Code, a community plugin by yuting0624, encodes a division of labor into your coding agent. Claude stays the conductor: requirements, architecture, the hard twenty percent, verification and review. The Antigravity CLI, Google's agy driving Gemini, becomes the executor: scaffolding, implementation, test generation, first-pass review, migrations and search, under a shared AGENTS.md so both sides follow the same conventions.
The README's justification is a quotable line about the current market: generation is solved, while verification, judgement and direction are the craft. The economics follow from the split, and the plugin is explicit that it is a community project, not affiliated with Google or Anthropic. Installation is the standard two-command plugin flow plus a setup command that verifies agy is installed and authenticated.
The measured part
Unusually for this category, the README ships numbers with their method. On a large ADK multi-agent build, same task and same model three ways: Claude solo at high effort cost an estimated 2.62 million cost-weighted tokens, solo at max effort 5.34 million, and the hybrid 1.91 million, with all three passing the same three-part adk evaluation. That is minus 27 percent versus solo-high and minus 64 percent versus solo-max, at equal quality, and the README notes the cheap Gemini work is not even counted in those figures.
The caveats are printed next to the numbers: costs are estimates, token counts are approximated, and rates live in prices.json, which you should set to your real Vertex rates before quoting anything. Savings scale with task size, and tiny one-off tasks are cheaper to just run on Claude. A full A/B document is committed in docs, and the break-even judgement is left to the user rather than automated away.
What the plugin adds beyond delegation
Several capabilities fill gaps in Claude's native toolkit. Media understanding delegates audio, video and images to Gemini's multimodal processing, returning a timestamped digest while the full transcript is written to a file, so an hour-long recording never lands in Claude's context. Research commands let agy do grounded web legwork while Claude verifies citations across at least two sources. A Cloud Run debugging command digests failing service logs and infers root causes, read-only by default.
Internal fan-out lets one delegation spawn agy's own subagents on the cheap side, each leaving a readable trajectory you can audit with agy-trace. A SessionStart hook injects the cost-aware routing policy automatically, and the delegate subagent does file writing on Gemini so Claude spends no tokens generating file contents. Both nudges are advisory by design: the break-even judgment stays with Claude, which the README backs with its own measurement that full auto-routing is a net loss below the break-even point.
Installing, and the platform fine print
Installation is the two-command plugin flow plus a health check:
/plugin marketplace add yuting0624/antigravity-for-claude-code
/plugin install antigravity@antigravity-for-claude-codethen /antigravity:setup verifies that agy is installed, authenticated and that its model list responds. Prerequisites are the Antigravity CLI and Claude Code themselves, with Vertex recommended on both sides for the same-bill savings.
The platform fine print matters: macOS, Linux and WSL are the supported targets for headless delegation, and native Windows through Git Bash or MSYS is explicitly not recommended, because agy can hang with a zero-byte log when run without a real console. The wrapper now bounds that hang with a wall-clock guard returning a clean TIMEOUT, and the doctor command distinguishes a hang from an auth failure, but the README's advice for reliable headless use is unambiguous: run from WSL, macOS or Linux.
The slash commands
The plugin registers a command family that appears natively in Claude Code's slash menu. Setup runs the health check. Delegate hands a subtask to agy under cost discipline with an optional flash or pro tier, then verifies the result. Review produces an independent cross-model review of the current diff, with an adversarial flag, and Claude reconciles the opinions.
Research orchestrates deep research with citation verification, media handles audio, video and images with the digest-to-file pattern, and a Cloud Run debugging command pairs agy's log digestion with Claude's root-cause inference. Background jobs let a long delegation run while you keep working, managed through status, result and cancel commands, with the documented caveat that one-shot headless sessions must delegate synchronously because there is no later turn to collect from. A migrate command moves an existing Claude Code setup, skills, memory, MCP servers, permissions, onto agy, dry-run by default with backup on apply and a reversal path.
Limits and positioning
The constraints are stated plainly. It is a community project without corporate affiliation, so both vendors can change the ground: the Antigravity CLI's version gates some features, and the Windows console issue shows how platform edges bite. The savings require the two-sided setup, Claude plus a Gemini-side account, and the estimates depend on the rates you enter. Three releases exist, the last push was on 2026-09-17, and a changelog and security policy are present, so maintenance is active.
Against plain Claude Code, the plugin is a cost and capability extension, not a replacement: Claude keeps verification and direction, which is where the README locates the craft. Against other delegation approaches, the differentiators are the measured results, the token-discipline hooks, and the audit trails. If your workload is large builds with separable bulk work, the hybrid math in the README is the argument; if it is small tasks, its own numbers say stay on Claude.
Editorial conclusion
Antigravity for Claude Code fits teams running large, separable builds through Claude Code who also have Gemini-side access, want cross-model verification and media understanding, and will set their real rates in prices.json before trusting the savings numbers. Skip it on native Windows, for tiny tasks where its own data says delegation loses money, or if adding a second vendor's CLI to your loop is more complexity than the 27 percent is worth. Verify on your workload: run one real task solo at high effort and once through the delegate command, and compare the dashboard costs and the adk-style quality checks before committing.
Frequently asked questions
Can Antigravity replace Claude Code?
Not in this design. The plugin keeps Claude as the conductor for requirements, architecture, verification and review, while the Antigravity CLI executes bulk implementation, tests and search, and the README's own measurements show the hybrid saving money at equal quality.
Is Google Antigravity like the Claude Code?
They have different strengths, which is why the plugin exists: the README assigns Claude the judgement and verification work and Antigravity's Gemini the cheap, fast throughput such as scaffolding, test generation and search.
Which platforms does Antigravity for Claude Code support?
macOS, Linux and WSL are supported for headless delegation. Native Windows through Git Bash or MSYS is not recommended because agy can hang without a real console, though the wrapper now bounds that with a timeout guard.
Community notes