Claude Code is one model.
Antigravity gives it a second opinion from another.
A single model agreeing with itself is not a review. The Antigravity skill wires Google Antigravity — the agy CLI, running Gemini 3.5 Flash with High Reasoning — into Claude Code as a callable second-opinion agent. When Claude is about to guess, push back, or commit to a design call, you can ask another model family to weigh in without leaving the session.
It is the Gemini sibling to the existing codex skill: same shape, different model behind it.
Use the skill when you want an independent take, not a second pass from the same model:
- a bug that survived Claude's fix
- an architecture decision where the trade-off matters more than the answer
- a stack trace nobody on this side of the model can read clearly
- a moment where you are about to accept a guess instead of a verification
Skip it for routine questions Claude can answer alone, and for anything that needs deep local codebase context agy does not have.
Trigger phrases inside Claude Code:
- "frag antigravity ..."
- "ask antigravity ..."
/antigravity ...
Claude Code invokes a thin wrapper script — agy.ps1 on Windows, agy.sh on macOS/Linux — which calls agy over a real PTY, captures the response, and returns it to the conversation. No extra terminal, no copy-paste between tools.
The first call performs a 30-second health check against agy and caches the result for 24 hours, so subsequent calls go straight to the answer.
- Claude Code with skills enabled (
~/.claude/skills/exists). - Antigravity CLI (
agy):- Windows:
irm https://antigravity.google/cli/install.ps1 | iex - macOS / Linux:
curl -fsSL https://antigravity.google/cli/install.sh | bash
- Windows:
- One-time interactive login in a normal terminal (not from inside Claude Code):
This kicks off the Google OAuth flow; the token lands in the system keyring.
agy
git clone https://github.com/offbyone1/antigravity-skill ~/.claude/skills/antigravityThat is the whole install. Claude Code picks the skill up automatically.
Inside Claude Code:
frag antigravity: is this migration safe under concurrent writes?
Or with the slash command:
/antigravity review this rate-limit code for race conditions
The skill structures the call, runs agy, and returns the response to the conversation. Treat the answer as a second opinion, not ground truth — verify against the actual code before acting on it.
You can also call the wrappers directly:
Windows (PowerShell)
& "$env:USERPROFILE\.claude\skills\antigravity\agy.ps1" -Prompt "your question"macOS / Linux (Bash)
bash "$HOME/.claude/skills/antigravity/agy.sh" "your question"Both accept an optional timeout (default 300 s):
& "$env:USERPROFILE\.claude\skills\antigravity\agy.ps1" -Prompt "..." -TimeoutSeconds 600bash "$HOME/.claude/skills/antigravity/agy.sh" "..." 600The wrappers exit cleanly with codes Claude Code can route on:
| Code | Meaning | Resolution |
|---|---|---|
| 0 | OK | Answer on stdout |
| 2 | agy not installed |
Run the install command above |
| 3 | agy installed, login missing or unresponsive |
Run agy interactively, complete OAuth |
| 4 | Timeout | Increase -TimeoutSeconds or shorten the prompt |
agy 1.0.0 has no clean headless mode. agy --print writes directly to the console buffer instead of stdout, which means a normal pipe redirect captures nothing. The wrappers work around this by allocating a real PTY:
- Windows (
agy.ps1): spawns a hidden PowerShell withStart-Transcript, then parses the transcript file. - macOS / Linux (
agy.sh): usesscript(1)to allocate a PTY and parses the resulting log.
The --dangerously-skip-permissions flag is set so tool approvals in print mode do not block the call.
- Default: Gemini 3.5 Flash (High Reasoning)
- No CLI switch for model selection —
agydecides
| File | Purpose |
|---|---|
SKILL.md |
Skill manifest Claude Code reads — when to trigger, how to call |
agy.ps1 |
Windows PowerShell wrapper |
agy.sh |
macOS / Linux Bash wrapper |
README.md |
This file |
LICENSE |
MIT |
MIT — see LICENSE.