skill
claude-antigravity-orchestrator
Claude orchestrates; the Antigravity CLI (agy) does the work — web search, downloads, scanning, validation, and first-draft code/UI — on a separate quota, so Claude spends far fewer tokens. Capability
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/shihabshahrier/claude-antigravity-orchestrator itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 23, 2026antigravity-delegate
Make Claude the orchestrator and the Antigravity CLI (agy) the worker. Offload
token-heavy, low-judgment grunt work — web search, downloads, scanning big logs/codebases,
running validators — to the local agy CLI in headless mode. agy eats the volume; Claude
gets back only the small distilled answer and keeps the reasoning.
Built for workflows like benchmarking a context-heavy system against competitors, where the
search / download / validate legwork would otherwise burn a pile of Claude Opus tokens while
agy sits idle.
agyis Google's official replacement for the Gemini CLI, which sunsets for individual / Google One / AI Pro / Ultra tiers on 2026-06-18. It authenticates via the system keyring and exposes Gemini 3.x, Claude, and open-weight models.
you ──▶ Claude (orchestrator, keeps judgment)
│ delegates closed, bulky subtask
▼
agy -p "<task>" --model "Gemini 3.5 Flash (Low)" (eats search/download/scan tokens)
│ returns ONLY the distilled answer (plain text)
▼
Claude reads the answer ──▶ decision
Install
# 1. Antigravity CLI (the worker)
curl -fsSL https://antigravity.google/cli/install.sh | bash
# 2. the skill
bash install.sh
Installs to every Agent-Skills path (Claude Code, Codex/Amp/Goose, opencode, Antigravity,
OpenClaw). Manual: copy skills/antigravity-delegate/ into your agent's skills dir.
Requirements
- Antigravity CLI (
agy) —curl -fsSL https://antigravity.google/cli/install.sh | bash - That's it —
agyprints plain text, so nojqneeded.
Name clash: the Antigravity IDE ships its own
agylauncher. If both are installed, keep the IDE launcher asagyand rename the CLI toagy-cli(mv ~/.local/bin/agy ~/.local/bin/agy-cli). The wrapper prefersagy-cli, thenagy, and honorsAGY_BIN=/full/path/to/cli.
Usage
/antigravity-delegate "<task>" [--model "Gemini 3.5 Flash (Low)"] [--mode search|read|write] [--out FILE]
Or call the helper directly:
skills/antigravity-delegate/scripts/agy-run.sh "<task>" --mode search --model "Gemini 3.5 Flash (Low)"
| Arg | Default | Meaning |
|---|---|---|
| task | required | one closed, self-contained instruction with a bounded deliverable |
--mode | search | search/read = read-only intent; write = auto-approve tools |
--model | "Gemini 3.5 Flash (Low)" | run agy models for the list; use a Pro/Thinking variant for reasoning |
--dir | — | scope tool access to a directory (--add-dir) |
--out | — | agy writes full output to FILE; stdout is a one-line summary |
--timeout | 5m | print-mode wait timeout |
Models on tap (via agy models): Gemini 3.5 Flash, Gemini 3.1 Pro, Claude Sonnet/Opus 4.6,
GPT-OSS 120B — pick per subtask.
When to delegate
agy runs on your Google One / Antigravity quota — separate from Claude's tokens — so
anything it does well is tokens Claude never spends. Two buckets:
| Delegate ✅ | Keep in Claude ❌ |
|---|---|
| web search → answer | architecture / trade-off decisions |
| download files/datasets | security-sensitive code or review |
| grep huge logs/codebases | shipping code with no review |
| run benchmark/validator → PASS/FAIL | synthesizing the final result |
| summarize a giant file | anything tiny (overhead not worth it) |
| frontend/UI scaffold, boilerplate, codemod (draft) | |
| generate unit tests (draft) |
Rule: delegate when the cost is volume or the value is a reviewable draft. Keep it when the
value is a final decision. A draft from agy is an input to Claude's judgment, not a
substitute — Claude reviews before anything ships.
Which model?
agy models lists what your account can use. Route by subtask (quote the exact name):
| Subtask | Model |
|---|---|
| search / download / scan / validate | Gemini 3.5 Flash (Low) — cheapest, fastest |
| frontend/UI drafts, code-gen, multimodal | Gemini 3.5 Flash (High) |
| deep reasoning / long-context | Gemini 3.1 Pro (High) |
| correctness-sensitive code, test gen | Claude Sonnet 4.6 (Thinking) |
| hardest reasoning / knowledge | Claude Opus 4.6 (Thinking) |
| open-weight fallback | GPT-OSS 120B (Medium) |
Cheapest-capable wins; escalate only when the worker must reason or the draft must be high quality.
How it works
scripts/agy-run.sh runs agy -p "<task>" --model "<name>" (adding
--dangerously-skip-permissions in write mode, --add-dir to scope), and prints agy's
plain-text answer to stdout. Exit codes pass through from agy; the wrapper adds 2 for bad
input and 1 for an empty response.
See skills/antigravity-delegate/references/agy-cli.md for the verified headless CLI reference.
Agent support
Ships Agent-Skills SKILL.md plus rule mirrors for Cursor, Windsurf, Cline, Copilot, and
Codex. Edit only skills/ — the rest mirror it.
Contributing
See CONTRIBUTING.md. License: MIT.