agent-autoloop: let your CLI agent work in the background and only ask when unsure

The biggest waste in CLI agents is the human watching the terminal: type a prompt, stare at it run, wait for questions — close the window and the work dies. agent-autoloop, open-sourced by KKKKhazix on November 3, 2026 (MIT, 1,277 stars in two days), flips that: a very lightweight automation hub lets the agent pick up tasks in the background and follow through, staying quiet until something genuinely needs your call, at which point an approval card lands in Slack, Discord, Telegram-style desktop notifications and Chinese IMs like Feishu, DingTalk or WeCom. No web dashboard, no cloud server, no open ports — one Mac or Linux box and one command.

Core features

• Zero-dependency kernel: task scheduling is a kernel with no external services — Node 24, Git and curl are the whole prerequisite list, and no GitHub account is needed.
• Plugin architecture: event sources, agents, notification channels and deploy pipelines are all pluggable; onboarding a new agent takes no SDK — any CLI tool that accepts a prompt and prints a final reply works.
• Waiting that costs nothing: when the agent ends a round with a question, it holds no process and burns no tokens, suspended until your decision arrives, then resumes from the breakpoint.
• Approval cards: decisions reach you as cards in the IM you already use; one click returns the verdict, and a single command shows overall progress.
• Ready-made templates: the repo ships scenario templates, and agent-autoloop try walks your own agent through a real ask-wait-decide-finish loop after install.

Typical use cases

• Overnight batches: queue code review, dependency bumps and doc updates for background agents, and start the morning with only the approval cards.
• Multi-agent pipelines: an event triggers one agent, its output hands off to the next, and the deploy plugin carries it to production.
• A personal automation hub: drop a task in one line and let the agent follow through, replacing ten terminal windows each doing its own thing.

Quick start

curl -fsSL https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.sh | sh
agent-autoloop try

macOS and Linux only; Node 24, Git, curl and any logged-in CLI agent are the prerequisites. The try command runs your agent through a full closed loop so you can see task-out, agent-working, suspended-question, decision, completion end to end.

Summary

agent-autoloop fits developers who already treat Claude Code or Codex as their main tool and want them off terminal duty; web panels, team permission management and multi-tenancy are explicitly out — single machine, single user, no UI is the design. Facts and limits: MIT, TypeScript, zero external dependencies; open-sourced November 3, 2026, so it is very early — expect the plugin interface and protocol to move, and trial it on low-stakes tasks first. It belongs to the class of tools that turn agents from sessions into staff: the scheduling kernel is small, and the value is in where it draws the plugin boundaries — it never touches your agent itself and never locks in a notification channel. For a different point on the same spectrum, see our writeup on deepseek-bot, where a team of bots dispatch work to each other in group chat rather than one agent looping in the background.