@loop.js/core

A thin, opinionated runtime for autonomous agent loops: state a Goal, and the engine drives Rounds — an Execute agent works in the work tree, then a separate, skeptical Verify agent (read-only by default) judges the result against the bar — until the Loop settles. A not-yet verdict must say why, and that reason feeds the next Round; a Goal that can never pass settles as impossible instead of burning budget.

Install

npm create @loop.js@latest my-loop   # scaffold a project: loop.config.ts + workspace/
cd my-loop && npm install

Or add the engine to an existing project: npm install @loop.js/core.

Agents run on the Claude Agent SDK — set ANTHROPIC_API_KEY.

Define

// loop.config.ts
import { Loop } from "@loop.js/core"

export default Loop.define({
  goal: "Build a playable 2D platformer in ./game.",
  limits: { rounds: 3, usd: 1 },
})

A prompt is a string, { file: "./goal.md" } (re-read fresh every Round — edit the bar mid-loop), or (ctx) => string.

Run

loop run          # drives Rounds until the Loop settles — or a guard fires; the exit code says which
loop status       # the Loop's Status so far — human-readable, or --json
loop cron add "0 8 * * *" --until settled   # a scheduler re-triggers it: crontab, launchd, schtasks, or Modal in the cloud

Or embedded — run claims the Lock and self-drives; iterating is a view, not a driver:

const run = loop.run()
for await (const event of run) { /* breaking out unsubscribes — it does not cancel */ }
const exit = await run.done()

Full docs: https://github.com/loop-js/loop.js · Apache-2.0