A manager for agentic loops. Schedule AI agents like cron jobs, watch them work in real terminals, and stop burning tokens on idle polling.
An agent that polls for its own work wastes tokens and context on every empty check — and its conversation eventually fills up and gets compacted. Looper inverts the loop: a cheap check script (and optionally a small classifier model) decides whether there is work, and only then a fresh agent session is started. No idle burn, no context rot, at most one agent per task, ever.
- Cron-style scheduling — wall-clock slots, timezones, schedule end dates; or manual tasks you fire on demand.
- Token savings by design — a shell check and an optional haiku-class classifier gate every run, so the expensive model only starts when there is real work to do.
- Fresh session every run — no context accumulation; the agent signals
looper-done, files its report, and the loop resumes. - Monitoring built in — per-task run log with results and reports, live status, and a real terminal tab you can watch and type into mid-run.
- System notifications — each task picks what it toasts, from errors-only to every run; clicking one lands on the live terminal or the run's log entry.
- Any harness, any environment — Claude Code, Codex, or a custom agent CLI; Looper on Windows driving agents inside WSL, or Linux/WSL directly.
- Easy task setup — an editor with Schedule / Check / Classifier / Agent tabs, templates, import/export; agents and scripts can register tasks by dropping a JSON file in the inbox.
IDLE ──schedule──▶ CHECKING ──act:false──▶ IDLE
CHECKING ──act:true──▶ [CLASSIFYING ──no──▶ IDLE] ──yes──▶ RUNNING ──done──▶ IDLE
- Check (optional) — your script runs in the task's directory and prints
one JSON line:
{"act": true, "summary": "3 new issues", "context": {…}}. - Classify (optional) — a cheap model gets the summary and answers
{act, reason}under a strict schema; its cost is recorded in the run log. - Agent — a fresh session of the task's harness starts in the task's
directory with your prompt (the check output templated in). It works in a
real terminal inside Looper and ends the run with
looper-done <status> "<headline>"plus a detailed report.
The check and the classifier are both optional — without them, every slot goes straight to the agent. The full contract (prompt templates, every way a run can end, held runs, idle detection) is in How a run works.
No compiler toolchain needed — the terminal component ships prebuilt binaries for Windows, Linux and macOS.
npm install
npm run dev # run from source (Electron, hot reload)
npm run package # or build the installer for your platform (release/)On first launch Looper creates a Local Shell environment pointed at the
claude CLI (plus a WSL bridge on Windows). Then:
- File → New Task… (Ctrl+N).
- On the General tab, name the task and pick its working directory; on the Agent tab, write the prompt. Optionally add a check command on the Check tab. Save.
- Select the task and press F5 (Run Now). Watch the agent work in the Terminal tab.
The Getting started guide walks through this with more detail, including the packaged Windows installer and notification setup.
The looper CLI and an inbox drop-folder let scripts — and other agents —
register and control tasks:
looper example > task.json # template task definition
looper add task.json # register it with the running app
looper run <taskId>
looper serve # run the engine headless, without the UISee CLI and automation for the full command set,
the inbox protocol, and a tour of the ~/looper data directory.
The full user guide lives in docs/:
- Getting started — install, first launch, first task.
- The task editor — every tab and field, templates, import/export, prompt template variables.
- How a run works — the check → classify → agent cycle and every way a run ends.
- Environments — local shell, WSL and Windows bridges, harnesses and their models, concurrency caps.
- Monitoring — terminal tab, run log, notifications, Rest Mode, tray.
- CLI and automation — the CLI, the inbox protocol, the data directory.
- Troubleshooting — common problems, current limitations.
npm run dev # Electron app with hot reload
npm run build:all # out/ (app) + out/cli/ (CLI + terminal worker)
npm test # engine unit tests
npm run typecheckThe engine has no Electron imports, so it also runs standalone
(looper serve). Packaging targets: npm run package:windows,
package:mac, package:linux.
PolyForm Strict 1.0.0 — free to use for noncommercial and personal purposes; no redistribution or derivative works.