keli-wen/agy-staff

What it claims

agy-staff runs Google's Antigravity CLI (agy, Gemini 3.8 Flash) as a worker for Claude Code, Codex, Pi and OpenCode. The main finding of Bench v1 is that the host matters. With agy, Claude Code's host spend fell on all three benchmarks (−16% to −50%). Codex's host spend rose on two: SWE-bench +86% and QnA +22%. The page attributes the gap to waiting behaviour. Claude Code waits in the background until the agy job ends. Codex polls, and every wake-up carries its full context. DeepSearchQA (research) is the only benchmark where agy cut spend on both hosts: Claude Code −50%, Codex −16%. "Host spend" means what the Claude or ChatGPT plan pays.

How it works

The design is bash plus skills, with no runtime library. A companion/ entrypoint handles modes and job commands, and templates/ holds shared prompts. The ask persona answers in the same call and has no tools. The other personas return a job id and a collection command such as wait <id> --timeout 10m. Expiring a wait leaves the worker running. observe reads a snapshot of recent activity, but only when the user asks about progress. cancel, continue and restart manage jobs. Execution deadlines default to 60 minutes and go up to 120.

The reviewer persona is prompt-only. agy gathers its own evidence (gh pr view, git diff), and no flag exists to pass it a diff. Under lead, the host delegates to staffer by default and keeps acceptance, integration and delivery for itself. Staffer also exposes agy's native tools, such as generate_image. On agy v1.1.15 this produced a 1024×1024 PNG in about 30 seconds.

Evidence

Bench v1 ran each task three ways: solo, with the host's own subagents, and with agy through lead. Tasks solved matched solo or dropped by one. Median time per task rose from 2 to 8 minutes solo to 13 to 26 minutes with agy. There was one run per task and setup, and the author calls it "a first baseline, not a verdict".

How to use it

Install agy and confirm with agy --version. Node.js is required.

curl -fsSL https://antigravity.google/cli/install.sh | bash

Claude Code:

claude plugin marketplace add keli-wen/agy-staff
claude plugin install agy@agy-staff

Codex:

codex plugin marketplace add https://github.com/keli-wen/agy-staff
codex plugin add agy@agy-staff

Pi: pi install git:github.com/keli-wen/agy-staff. OpenCode V1 (verified on 1.18.34):

opencode plugin 'agy-staff@git+https://github.com/keli-wen/agy-staff.git' --global

Restart the harness, then smoke-test with /agy:ask reply with OK (Codex uses $agy:ask, Pi uses /skill:agy-ask, OpenCode uses /agy-ask). Upgrades are manual. Claude Code and Codex cache per version directory, so an upgrade lands only when the plugin version changes. To upgrade OpenCode, pin a newer tag or commit with --force. The licence is MIT.

Limits

staffer, researcher, reviewer and implementer run unrestricted by default. agy can read the repo, run commands, edit files, and commit, push or open PRs if the task asks for it. Guards such as warning agy about pre-existing dirty files are prompt instructions and "do not provide permission isolation". Optional hardening uses --restricted per run or setup --restrict review,research per repo. Its allowlist is prefix-matched and applies machine-wide, and restricted runs can return less output. OpenCode V2 is not supported. Codex users should expect higher host spend outside research tasks.