# Antigravity for Claude Code keeps the judgement and hands Gemini the bulk work > A Claude Code plugin that calls the Antigravity CLI as a sub-agent for scaffolding, implementation, test generation and search, while Claude holds requirements, verification and review. Its own numbers are labelled estimates, and native Windows is the platform it tells you not to use. **yuting0624/antigravity-for-claude-code** — 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. - Repository: https://github.com/yuting0624/antigravity-for-claude-code - Stars: 367 · Forks: 37 - Language: Shell - License: MIT - Published: 2026-09-18 · Updated: 2026-09-18 · Language: en - Canonical page: https://hysenlabs.com/projects/yuting0624-antigravity-for-claude-code ## Two roles, one shared AGENTS.md The split is explicit and asymmetric. Claude is the conductor and owns requirements, architecture, the difficult remainder of the work, verification and review. The executor is the Antigravity CLI, agy, backed by Gemini, and it owns scaffolding, implementation, test generation and search. The stated strengths follow from that: judgement on one side, cheap fast throughput on the other. ``` you → Claude Code (conduct: design / verify / review) └── agy → Gemini (execute: implement / test / search) ``` Both sides read a shared AGENTS.md, which is what stops the executor from drifting once a task leaves Claude. The plugin also adds capabilities Claude Code has no native equivalent for, including live Google and web search, Vertex AI Search over internal data, deep research and Cloud Logging, with Claude re-checking what comes back. ## Installation is a marketplace entry, then a setup check There is no package to build and no script to fetch by hand. Three lines inside Claude Code do it: ``` /plugin marketplace add yuting0624/antigravity-for-claude-code /plugin install antigravity@antigravity-for-claude-code /antigravity:setup # verifies agy is installed + authenticated ``` The setup command is the one that matters. It checks that the Antigravity CLI is present and authenticated, which agy models confirms by listing the Gemini models available to it. Claude Code itself is the other prerequisite, and the plugin notes that running Claude Code on Vertex as well is what puts both sides on the same bill, which is where the cost claim comes from. ## The savings table is labelled an estimate twice The measured section covers one large ADK multi-agent build with adk eval, run three ways on the same task with the same model. Frontier cost is given as 2.62M for Claude solo at high, 5.34M for solo at max and 1.91M for the hybrid arrangement, with quality at 3 out of 3 in all three columns. The stated result is 27 percent below solo at high and 64 percent below solo at max at equal quality, and the Gemini work is not counted in that. Two caveats sit next to the table. The figures are estimates with approximated token counts, and the rates live in prices.json. The instruction is to set real Vertex rates there before quoting anything. The full comparison is kept separately in docs/AB-RESULTS.md, and the plugin states that savings scale with task size, so a tiny one-off task costs more delegated than run outright. ## Automatic routing is advisory and stops below break-even A SessionStart hook injects the cost-aware routing policy into the session, and the toggle sits in plugin settings. Alongside it, the antigravity-delegate subagent does the file writing on the Gemini side, so Claude spends no tokens generating file contents at all. Delegation is not tied to a slash command. The subagent is picked up proactively when a request looks like bulk work, and a prompt-level nudge flags candidates, toggled through delegation_nudge. Both mechanisms are described as advisory, and the reason given is a measurement: full automatic routing comes out as a net loss below the break-even point, so the break-even call stays with Claude. That is why the delegate command still takes an explicit tier flag, flash or pro. ## Background jobs have no second turn to be collected in claude -p Long delegations can be fired and left, then collected through the status, result and cancel commands. That model assumes a person is there, and the documentation is blunt about the alternative: in headless claude -p, delegate synchronously, because a one-shot invocation has no later turn in which to collect anything. The same advice applies to what the executor does next. Internal fan-out is available, where agy spawns its own subagents on the cheap side, using dynamic define_subagent on agy 1.0.16 and later, or TypeName "self" with a Role on any version. Each leaves a trajectory that agy-trace reads back for audit, which is the only handle on what happened while the job was running. ## A one-hour recording arrives as a digest, not as context The media command handles audio, video and images by delegating perception to Gemini, which is multimodal on its own. That removes the need for a local ffmpeg and Whisper stack on the Claude side, and it changes what reaches the context window: you get a timestamped digest back, while the full transcript is written to a file, so a long recording does not sit in the session. The other commands follow the same shape of delegating the heavy legwork and keeping the judgement local. Research has agy do grounded web work while Claude verifies citations across at least two sources. Review produces an independent cross-model opinion on the current diff that Claude then reconciles, and it takes an adversarial flag. Cloud Run debugging digests error logs for a root-cause inference, staying read-only unless the apply flag is passed, in which case the write goes to a branch. ## agy -p hangs without a real console, so Windows Git Bash is ruled out macOS, Linux and WSL are the supported targets for headless delegation. Native Windows under Git Bash or MSYS is explicitly not recommended, because agy -p can hang with a zero-byte log when it runs without a real console, which is the ConPTY problem tracked in issue 6. The wrapper bounds the damage rather than fixing the cause. A wall-clock guard built on GNU timeout, or gtimeout, returns a clean TIMEOUT instead of hanging, and the doctor command separates a hang from an authentication failure so the two are not confused during setup. Even with that guard in place, the recommendation is to run from WSL, macOS or Linux when the work is unattended. ## migrate reads your skills in place instead of copying them The migrate command moves an existing Claude Code setup onto agy: skills, CLAUDE.md, memory, MCP, plugins and permissions. It prints a plan by default, backs up before applying, and reverses with uninstall plus apply. ``` /antigravity:migrate # see the plan /antigravity:migrate --apply # global assets /antigravity:migrate --apply --include-repos # also AGENTS.md + per-repo memory ``` The asset mapping avoids duplication where it can. Skills are read in place through a skills.json entry rather than copied, so one edit serves both tools. CLAUDE.md becomes an AGENTS.md symlink beside it, and installed plugins land in the Gemini config plugin directory through the native importer with its output repaired. Your home directory is never written to. The table describing auto-memory rules stops partway through the per-repository row, so what follows that row is not visible here. ## Conclusion This fits large builds where token volume dominates, since the reported saving comes from a multi-agent ADK task and the plugin says small one-off jobs are cheaper to run on Claude directly. It does not fit a Windows Git Bash setup, where the wrapper's wall-clock guard exists only because agy -p hangs without a real console, so use WSL. Two things to check before quoting any figure: put your real Vertex rates into prices.json, because the headline numbers are approximations, and confirm the break-even yourself, since the plugin deliberately leaves automatic routing off below that point. It is a community project with no affiliation to Google or Anthropic. ## FAQ ### Does Antigravity replace Claude Code in antigravity-for-claude-code? No. Claude stays the conductor and keeps requirements, architecture, verification and review, while the Gemini-backed agy CLI handles scaffolding, implementation, test generation and search. ### Which platforms does antigravity-for-claude-code support for headless delegation? macOS, Linux and WSL. Native Windows under Git Bash or MSYS is not recommended because agy -p can hang with a zero-byte log when it runs without a real console, tracked as issue 6. ### How does antigravity-for-claude-code report its cost savings? On one large ADK multi-agent build with adk eval, frontier cost is listed as 2.62M for solo at high, 5.34M for solo at max and 1.91M for the hybrid, with quality at 3 out of 3 in all three. The plugin labels these figures estimates. ### What does the antigravity setup slash command check? It is a health check that confirms the Antigravity CLI is installed and authenticated and that the scripts are ready. agy models lists the Gemini models available to the CLI. ### Does the migrate command in antigravity-for-claude-code copy my skills? No. Skills under ~/.claude/skills are read in place through a skills.json entry rather than copied, so one edit serves both tools, and CLAUDE.md becomes an AGENTS.md symlink beside it. Your home directory is never written to. ## Sources - [Issues](https://github.com/yuting0624/antigravity-for-claude-code/issues) - [License: MIT](https://github.com/yuting0624/antigravity-for-claude-code/blob/master/LICENSE) - [README](https://github.com/yuting0624/antigravity-for-claude-code/blob/master/README.md) - [Releases](https://github.com/yuting0624/antigravity-for-claude-code/releases) - [yuting0624/antigravity-for-claude-code on GitHub](https://github.com/yuting0624/antigravity-for-claude-code) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/yuting0624-antigravity-for-claude-code