Agent Autoloop: Let the command-line Agent handle tasks in the background on its own, and only come to you for decisions when it's unsure.Agent Autoloop:让命令行 Agent 后台自己接活,拿不准才来找你拍板
The biggest waste of a command-line agent is having someone man the terminal: if you walk away, it stops. Agent Autoloop is a Chinese open-source project that was only created on 3 November and reached 5,909 stars in just two weeks – an automation middleware with zero external dependencies. It turns event sources, agents, and notification channels into pluggable plugins. Tasks queue up and follow up on their own, and only when a decision is needed will a card be pushed to Lark, DingTalk, WeChat Work, or Slack to notify you.
命令行 Agent 最大的浪费是人守着终端:你走开,它就停。Agent Autoloop 是 11 月 3 日才创建、两周冲到 5909 星的国产开源项目——一个零外部依赖的自动化中台,把事件来源、Agent、通知通道全部做成可插拔插件,任务自己排队跟进,只有需要拍板时才把卡片推到飞书、钉钉、企微或 Slack 找你。
What Problems Does It Solve?它解决什么问题
Normally, command-line agents like Claude Code and Codex are used "with a person watching the terminal": you type a sentence, watch it run, and wait for it to ask what to do next. As soon as you close the window or walk away, things stop. Agent Autoloop flips this interaction around – once a task is issued, the agent works on its own in the background, without taking up a process or burning tokens, and only reaches out to you when it encounters something it's unsure about.
The architecture is "kernel + plugins": a task scheduling kernel with zero external dependencies, with event sources, Agent integration, notification channels, and deployment pipelines all being pluggable. A new Agent can be connected with a single line of configuration, without the need for an SDK or modification of the source code – any command-line tool that "can receive prompts and print the last response" can be connected.
Zero barriers to deployment: It can run on a single computer with a single command, without the need to purchase a cloud server or open a public network port. There is no backend interface, and you can see the overall progress by typing a single line of code, agent-autoloop status, which includes the number of rounds, tokens, and cost statistics for today and the past 7 days.
平时用 Claude Code、Codex 这类命令行 Agent 是「人守着终端用」:输一句话、盯着跑、等它问下一步;一旦关窗口或人走开,事情就断了。Agent Autoloop 把这套交互倒过来——任务发出后 Agent 在后台自己干活,不占进程、不烧 token,遇到拿不准的事才主动找你。
架构上是「内核 + 插件」:零外部依赖的任务调度内核,事件来源、Agent 接入、通知通道、上线流水线全部可插拔。一行配置就能接新 Agent,不需要 SDK 也不改源码——任何「能接收提示词并打印最后一条回复」的命令行工具都能接进来。
部署零门槛:一台电脑一条命令就能跑,不用买云服务器、不开公网端口;没有后台界面,全局进展敲一行 agent-autoloop status 就能看到,包括今天和最近 7 天的轮数、token 和费用统计。
How does something turn?一件事怎么转起来
The process consists of four steps (corresponding to the numbers in the architecture diagram): An event comes in → The kernel dispatches it to the Agent → If a decision is needed after the Agent completes one round, the "Waiting for your decision" card is pushed to your Lark/DingTalk/WeCom/Slack/Discord or as a desktop notification → You reply with the number, and the kernel then completes the task.
Built-in test run: After installation, run agent-autoloop try, and it will go through the process of "asking a question → exiting and waiting for a response → making a decision → then completing the task". For non-interactive terminals, use agent-autoloop answer 1 2 to answer the card, then use status to view the result.
The project was created on 03/11/2026 by KKKKhazix, with an MIT open-source licence. The current version, v1.9.3, is still undergoing frequent iterations (commits were still being pushed at 00:00 on 19 November). The repository includes "ready-to-use scenario templates" and documentation for a plugin system that can be replaced in all four directions.
流程分四步(对应架构图编号):事件进来 → 内核派发给 Agent → Agent 干完一轮若需要决策,就把「等你决定」卡片推到你的飞书/钉钉/企微/Slack/Discord 或桌面通知 → 你回复序号,内核接着把活干完。
上手自带试跑:安装后执行 agent-autoloop try,会真实走一遍「提问 → 退出等回复 → 拍板 → 接着干完」。非交互终端用 agent-autoloop answer 1 2 回答卡片,再用 status 看结果。
项目 2026-11-03 创建,作者 KKKKhazix,MIT 开源,当前 v1.9.3 仍在高频迭代(11 月 19 日凌晨还在推提交);仓库里附了「拿来即用的场景模板」和四个方向都能替换的插件体系文档。
Download & Install下载与安装
Click the button above to download agent-autoloop-main.zip (0.69MB, the complete source code from the GitHub repository, including the kernel, all plugins, and SKILL). After unzipping, give the installation instructions INSTALL.md to your Agent and let it install itself – this is the officially recommended approach.
Official one-line installation script: For Windows PowerShell, run irm https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.ps1 | iex; for macOS/Linux, use curl -fsSL .../install.sh | sh. All installations are of the latest Release, which has passed all checks.
System requirements: Windows, macOS, or Linux; Node 24 or later; Git; and a logged-in command-line agent. macOS/Linux also require curl. A GitHub account is not required at any stage.
点上方按钮下载 agent-autoloop-main.zip(0.69MB,GitHub 仓库完整源码,含内核、全部插件和 SKILL)。解压后把安装说明 INSTALL.md 丢给你的 Agent,让它自己装是官方推荐姿势。
官方一行安装脚本:Windows PowerShell 执行 irm https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.ps1 | iex;macOS/Linux 用 curl -fsSL .../install.sh | sh。安装的都是通过全部检查的最新 Release。
环境要求:Windows、macOS、Linux 均可,Node 24 或更新、Git,以及一个已登录的命令行 Agent;macOS/Linux 还需要 curl,全程不需要 GitHub 账号。
Note小提示
Don't assign important tasks to it right away: first, use agent-autoloop try to familiarise yourself with the card approval process, and then start queuing up low-risk tasks such as "Organise the inbox at a set time every day". It is recommended that card notifications be linked to at least one channel that you frequently check on your phone; otherwise, the backlog of tasks awaiting approval will get stuck. The Agent Autoloop source code package that accompanies this article can be found via the download button above.
别一上来就把重要任务交给它:先用 agent-autoloop try 熟悉卡片拍板流程,再从「每天定点整理收件箱」这类低风险任务开始排队。卡片通知建议至少配一个你手机常看的渠道,否则拍板积压任务会卡住。本文配套 Agent Autoloop 源码包见上方下载按钮。
⬇ Download · 点击下载:Agent Autoloop v1.9.3 完整源码包(含内核与全部插件)
MIT · GitHub 仓库 zip · 需 Node 24+,解压后按 INSTALL.md 安装;也可用官方脚本一行安装
阅读与点赞数据保存在你的浏览器本地,欢迎留下你的想法。
今日正能量学一点,用一点;今天种下的种子,会长成明天的能力。去免费下载专区 →广告
评论 文明发言,让讨论更有价值