Turn Claude Code or Codex into an orchestrator — hand heavy jobs to Antigravity sub-agents and keep only the answers.
Created by Zarfiks
A skill: one SKILL.md file plus a few small scripts. It teaches your AI coding
agent (Claude Code or Codex) to pass work to the
Antigravity CLI (agy).
Your agent becomes the boss. agy workers are the helpers. A helper
reads the files, does the job and returns a short answer. The boss never loads
those files, so its context stays small and it keeps working on your task.
your repo ── Claude Code / Codex (boss) keeps coding here
│
├── copy 1 ── agy helper: review ─┐
├── copy 2 ── agy helper: security check ─┼──► short answers ──► boss decides
└── copy 3 ── agy helper: fix src/billing ─┘──► diff ──► boss checks, merges
Every helper works in its own copy of the repo, so helpers never edit the files the boss is editing.
Good for: reading a big log, checking the same thing in ten modules, review and security passes while the boss codes, a second opinion from another model, small independent fixes.
Not for: small jobs. Every helper costs about 12 000 tokens and a few
seconds. Two greps are faster.
You need: the Antigravity CLI (agy) signed in,
git, Python 3, and bash (on Windows, Git Bash).
git clone https://github.com/Zarfiks/Antigravity-Slavery.git
cd Antigravity-Slavery
./install.sh # Claude Code
./install.sh --codex # CodexRestart your agent. Done.
Just ask your agent:
"Use agy helpers to find every place we read config in this repo."
"Get a second opinion from opus on this function."
"Have an agy helper fix this bug in a worktree."
The agent picks the model and the number of helpers by itself.
| Name | Model | For |
|---|---|---|
gemini-high |
Gemini 3.8 Flash High | hard questions |
gemini-medium |
Gemini 3.8 Flash Medium | normal work |
gemini-low |
Gemini 3.8 Flash Low | simple, clear tasks |
opus |
Claude Opus 4.6 | second opinion |
sonnet |
Claude Sonnet 4.6 | cheaper second opinion, text |
gpt-oss |
GPT-OSS 120B | quick yes/no checks |
The names are fixed; the models behind them are found automatically from
agy models, newest first. If a model is busy, the next one is used.
# ask one helper
scripts/agy-slave.sh gemini-medium "Explain src/net/client.py" ./repo
# let a helper edit code: only src/billing/, then run the tests
scripts/agy-slave.sh -w -o src/billing/ -v "npm test" gemini-high "Fix the rounding bug" ./repo
scripts/agy-merge.sh /tmp/agy-worktrees/repo-... # path is printed; brings the change back
# a queue of helpers: one task per line, at most 3 at once, 1 of them writing
scripts/agy-fanout.sh -j 3 --max-write 1 -m notes.md gemini-medium ./repo examples/tasks.txt# two model families review the same code; you get what they agree on
scripts/agy-consensus.sh "Review src/auth/ for security bugs" ./repo
# which model is behind each name right now; what the helpers cost so far
scripts/agy-models.sh
scripts/agy-cost.sh --since 2026-09-23Put your project's checks in .agy-verify (one command per line, e.g.
npm run lint, npx tsc --noEmit, npm test) and every editing helper is
checked with them automatically.
On Windows PowerShell: scripts\agy.ps1 slave gemini-medium "..." C:\code\repo
(also fanout, consensus, merge, models, cost).
| Flag | What it does |
|---|---|
-w |
helper may edit; its changes wait in a snapshot for agy-merge.sh |
-o PATHS |
files the helper may change, e.g. src/auth/,docs/auth.md |
-v "CMD" |
a check after the helper (repeat: lint, typecheck, tests, build) |
-s NAME |
answer as JSON: findings, list, verdict or your schema file |
-m FILE |
shared notes between helpers |
-f |
allow shell commands (use with care) |
-c ID |
continue an earlier helper |
-t SEC |
time limit |
All options: scripts/agy-slave.sh -h. Details and fixes:
references/.
- Helpers never work in your folder. Each one gets its own copy of the
repo (a
git worktreewith your uncommitted files too). You and five helpers can work at the same time without stepping on each other. - Read helpers' copies are thrown away. Write helpers' changes come back only
when you run
agy-merge.sh— after you looked at the diff. - Each write helper owns its files (
-o). Two helpers on the same file are refused up front, or flagged asCONFLICTafter a fan-out. agy-merge.shmerges line by line if you changed the same file meanwhile; real clashes get conflict markers. Nothing is committed for you.- By default helpers cannot run shell commands.
-wand-fallow it. Do not use-fon code you do not trust. - Your prompts and files go to the model provider. Do not give helpers secrets.
Not affiliated with Google, Anthropic or OpenAI. All names are trademarks of their owners.