Prompt & Done · AUTOLOOP
Let Claude Code work in the background and ping you only when a decision is yours
Agent Autoloop (free, MIT) turns Claude Code or Codex into a background loop: the agent works one round, exits, and sends you a card only when it needs your call. While it waits, no agent process runs and no tokens are spent; your answer resumes the same conversation.
Tested 7 Nov 2026 on a Mac with v1.6.2: a focus-timer web app + tests ran 6 min 42 s unattended, asked one question (25 or 50 minutes?), waited 1 min 55 s with zero agent processes, then finished 25 s after we answered. 32 tests passing.
1 Install (macOS / Linux)
You need Node 24 or newer, Git and a signed-in claude (or codex) CLI.
curl -fsSL https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.sh | shWindows PowerShell: irm https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.ps1 | iex. The installer sets up a small background hub; desktop notifications are on by default (Slack, Discord, Feishu and others are optional).
2 Try the full loop in a minute
agent-autoloop tryThe agent asks you one test question and exits. Answer it with agent-autoloop answer 1 2, then agent-autoloop status shows it picked the same session back up and finished. Heads-up: CLI and notification text are still in Chinese (English README; the agent writes its questions in your prompt's language; ours came back in English).
3 Give it a real job (the rule we used)
Add a rule to ~/.agent-autoloop/rules.json, then send the event with agent-autoloop emit build.start:
{ "name": "build-app", "when": { "source": "local", "types": ["build.*"] },
"to": "claude", "cwd": "~/code/nightshift",
"prompt": "Build a single-page focus timer ... cover the logic with node:test tests and run them until they pass. Commit your work. Make every technical and design call yourself. One thing is my call: whether a session defaults to 25 or 50 minutes." }Tell it what's yours to decide; it decides the rest. With no reserved decision, our first run finished in 8 minutes without a single ping. Rules can also run on a schedule ("cron": "0 9 * * *") or on GitHub issues.
4 Before you leave it running
Model usage is billed to your own Claude or OpenAI account. For automated workloads, providers recommend API keys over personal subscriptions (see the README's "Usage and provider terms"). Agents run with the access you give them in rules.json (default: edit); use a throwaway folder first.