中文版

agent-autoloop（GitHub：https://github.com/KKKKhazix/agent-autoloop ，2026-11-03 创建，今晨抓取时已 4,103★）是一个极轻量的自动化中台：让 Claude Code、Codex 这类命令行 Agent 在后台自己接活、自己跟进到底，平时不烦你，只有遇到拿不准的事才发一张卡片找你拍板。

README 点出的问题很实在：平时用命令行 Agent，大多是“人守着终端用”——你输一句话，盯着它跑，等它问你下一步；窗口一关、人一走开，事情就断了。agent-autoloop 把任务调度做成一个零外部依赖的内核，再把事件来源、Agent、通知通道、甚至代码上线流水线全部拆成可随时插拔的插件。一台电脑、一条命令就能跑，装在 Windows、Mac 或 Linux 上，不用买云服务器、不用开放公网端口；一行配置就能接新的 Agent，不需要 SDK、也不用改 Agent 源码，能接收提示词并打印最后一条回复的命令行工具都能接。

它没有后台界面：需要你拍板时把卡片推到飞书、钉钉、企微、Slack、Discord 或桌面通知；想看全局进展，敲一行命令就行。仓库里还给了拿来即用的场景模板和四方向的插件体系。安装需要 Node 24 或更新、Git 和一个已经登录的命令行 Agent，macOS/Linux 再加一个 curl；装好后跑 agent-autoloop try 就能真实走一遍“提问 → 退出等回复 → 拍板 → 接着干完”的完整流程。

为什么值得了解

• 它补上了命令行 Agent 最缺的一块：从“人守着用”变成“后台自己跑”，而且不绑云、不开端口，一台自己的电脑就是全部基础设施
• 插件化做得很彻底：事件源、Agent、通知、上线流水线四个方向都能自由替换，今天接 Claude Code，明天换 Codex 也不用重搭
• 中文作者、中文 README，场景模板直接按国内常用工具（飞书、钉钉、企微）来写，对国内用户几乎零上手成本；作者上一款 AIHOT 也是同一风格的热门项目

English Version

agent-autoloop (GitHub: https://github.com/KKKKhazix/agent-autoloop , created 2026-11-03, 4,103 stars at this morning’s fetch) is an extremely lightweight automation hub: it lets command-line agents like Claude Code and Codex take on work in the background and follow it through on their own, leaving you alone unless something needs your call — then it sends you a card to decide.

The README names a very real problem: using a command-line agent today mostly means “a human babysitting the terminal” — you type a sentence, watch it run, wait for it to ask what’s next; close the window or walk away and the work stops. agent-autoloop turns task scheduling into a kernel with zero external dependencies, and splits event sources, agents, notification channels and even the code-shipping pipeline into plugins you can swap at any time. One computer and one command is all it takes: install it on Windows, Mac or Linux, no cloud server to buy, no public port to open; one line of config connects a new agent, no SDK and no agent source changes — any CLI tool that takes a prompt and prints its final reply can plug in.

There is no dashboard: when your decision is needed, a card is pushed to Feishu, DingTalk, WeCom, Slack, Discord or a desktop notification; to see overall progress, run one command. The repo also ships ready-to-use scenario templates and a plugin system covering four directions. Installation needs Node 24 or newer, Git and an already-logged-in command-line agent, plus curl on macOS/Linux; after installing, agent-autoloop try walks you through a real “ask → wait for the reply → decide → finish the job” cycle.

Why it’s worth knowing

• It fills the biggest gap in command-line agents: going from “a human babysitting the terminal” to “running itself in the background”, without any cloud or open ports — one of your own computers is the whole infrastructure
• The plugin design is thorough: event sources, agents, notifications and the shipping pipeline can all be swapped freely, so switching from Claude Code to Codex tomorrow doesn’t mean rebuilding
• A Chinese author with a Chinese README, and scenario templates built around the tools people actually use in China (Feishu, DingTalk, WeCom), so the onboarding cost for domestic users is close to zero; the author’s previous project AIHOT was a hit in the same style