β†— Evidence-first loop engineering

Done means proven.
Not just said.

Type loopy <task> β€” an agent does the work, proves every piece with a real artifact, and only then says it's done. Like Luffy reaching the goal: it stretches as far as it takes.

Runs on Codex Β· Claude Code coming soon Β· zero runtime dependencies

$ loopy fix the failing login test and prove it
β€Ί plan ……… 3 steps
β€Ί proof …… .superloopy/evidence/login-test.txt
β€Ί gate …… quality check passed
βœ” done β€” backed by real evidence

Why Superloopy

"Done" should mean more than a confident sentence.

Superloopy is for agent work where the final gate has to be real. Every pass leaves proof behind.

🧾

Evidence-first

Every pass points at a real artifact under .superloopy/evidence/. No proof, no "done."

πŸͺΆ

Lightweight by default

One small CLI, repo-local state, zero runtime dependencies. Just Node β‰₯ 20.

🀝

Agent-friendly

Skills, hooks, and optional crew lanes guide Codex β€” and soon Claude Code β€” without ever hiding the final gate.

See it in action

Same prompt. A different finish.

One prompt for a fruit-commerce landing page, the same model both times β€” GPT-5.5, xhigh reasoning β€” run once with Superloopy, once without. The only thing that changed is Loopy.

01 The prompt Β· given to both
Build a clean, Stripe-brand-style landing page that introduces and sells beautiful, tasty-looking fruit. Shoppers should be able to add realistic, delicious-looking fruit to a cart and check out, and clicking a fruit should reveal its details.
Without Superloopy
Fruit demo built without Superloopy β€” live Open live demo β†—
Same prompt, single pass β€” no DESIGN.md gate, no anti-slop pass, no visual-QA evidence. Shipped on the first try.
πŸŒ€ With Superloopy
Loopy Fruits β€” live demo Open live demo β†—
Built by the frontend skill: a DESIGN.md token contract, anti-slop pre-flight, and a real-browser visual-QA gate. Cart, product detail & checkout all work.

Skills

A small command layer. Four specialist workflows on top.

Skills carry the workflow: when to use it, what to inspect, and what proof to leave behind. Every one finishes on the same evidence spine.

πŸ”

loop

loopy <task> (or 루피) runs a full evidence loop β€” plan, guided actions, command-backed proof, a quality gate, and a final report.

πŸ”¬

research

loopy research fans parallel read-only lanes over code, web, docs & OSS, chases every lead, verifies by running code, and synthesizes a cited report.

🎨

frontend

Auto-activates on any UI work β€” enforces a DESIGN.md token contract, bans AI-slop defaults, and gates on real-browser visual QA.

🧬

clone

loopy clone rebuilds an authorized site pixel-close β€” captures, tokens, assets, build, and visual QA.

The crew

Summon the crew for bigger work.

Six optional subagents, each owns one lane. loopy team <task> fans the work out in parallel β€” and still proves every piece before it calls it done.

franky
frankybuilds it
zoro
zororeviews it
usopp
usopptests it
jinbe
jinbegates it
robin
robinaudits it
nami
namifinds it

Install

Install on Codex. Then just type loopy.

Dependency-free β€” just Node.js β‰₯ 20 (Codex also needs Codex CLI β‰₯ 0.131.0).

Codex
$ codex plugin marketplace add https://github.com/beefiker/superloopy
$ codex plugin add superloopy@beefiker
Claude CodeComing soon
Native Claude Code plugin is on the way β€” the same loop, the same evidence gates.

Restart (or reload plugins) and approve the hooks when prompted. Then just type loopy <task> β€” or the Korean alias 루피. Verify with superloopy doctor.