AI资讯大全AIPPXP.CN搜索 ↗
图文教程 · 手把手

A Step-by-Step Guide to Setting Up Agent Autoloop: Let Claude Code Handle Tasks in the Background手把手教你装好 Agent Autoloop:让 Claude Code 在后台自己接活

In the previous article, we introduced Agent Autoloop, the "Agent Automation Mid-End". In this article, we'll guide you through setting it up from scratch and going through the complete closed loop of "Assign Task → Run in the Background → Approve Card → Complete Task". The official source code package released today will be used throughout the process, and it is compatible with Windows, macOS, and Linux.

上一篇我们介绍了 Agent Autoloop 这台「Agent 自动化中台」,这篇带你从零装好它,并真实走一遍「派活 → 后台跑 → 卡片拍板 → 干完收工」的完整闭环。全程用今天入库的官方源码包,Windows、macOS、Linux 通用。

2026-11-19 更新 · 免费
手把手教你装好 Agent Autoloop:让 Claude Code 在后台自己接活
1

Step 1: Prepare the environment and install第 1 步:准备环境并安装

First, confirm three things: Node 24 or later (check with node -v), Git, and a logged-in command-line Agent (this article uses Claude Code as an example). Windows users are recommended to operate in PowerShell.

Click the button above to download agent-autoloop-main.zip and extract it to any directory, such as C:\tools\agent-autoloop. Give the INSTALL.md file inside to your Agent and say "Install me according to this document". This is the officially recommended installation method.

Don't want the Agent to do it for you? On Windows, run irm https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.ps1 | iex in PowerShell; on macOS/Linux, run curl -fsSL https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.sh | sh. Both will install the latest Release that has passed inspection.

中文

先确认三样东西:Node 24 或更新(node -v 查看)、Git、以及一个已登录的命令行 Agent(本文以 Claude Code 为例)。Windows 用户建议在 PowerShell 里操作。

点上方按钮下载 agent-autoloop-main.zip,解压到任意目录,比如 C:\tools\agent-autoloop。把里面的 INSTALL.md 交给你的 Agent 并说「按这个文档把我装好」,是官方推荐的安装姿势。

不想让 Agent 代劳?Windows 在 PowerShell 执行 irm https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.ps1 | iex;macOS/Linux 执行 curl -fsSL https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.sh | sh,装的同样是最新通过检查的 Release。

第 1 步:准备环境并安装
2

Step 2: Play an official trial game to understand the bidding mechanism第 2 步:跑一次官方试玩,理解拍板机制

After the installation is complete, run agent-autoloop try. It will go through a complete closed loop in real time: The agent starts working → It finishes this round and asks you a question (at this point, it does not take up a process or use tokens) → The terminal displays the "[Waiting for your decision]" card, asking you to choose blue or green.

Directly enter the number and press Enter in an interactive terminal; if it is a non-interactive terminal, follow the prompt to execute agent-autoloop answer 1 2 to select "Green".

After answering, run agent-autoloop status: You can see that it continues from the previous conversation and displays "You chose green, the test run was successful", as well as statistics on the number of rounds, tokens, and costs for today and the past 7 days. Completing this step indicates that the kernel and your Agent links are all functioning normally.

中文

安装完成后执行 agent-autoloop try。它会真实走一遍完整闭环:Agent 开始干活 → 结束这一轮并向你提问(此时它不占进程、不花 token)→ 终端出现「【等你决定】」卡片,问你选蓝色还是绿色。

交互终端里直接输数字回车;如果是非交互终端,按提示执行 agent-autoloop answer 1 2 来选择「绿色」。

答完运行 agent-autoloop status:能看到它接上刚才的会话、显示「你选了绿色,试跑通过」,以及今天和最近 7 天的轮数、token 与费用统计。走通这一步,说明内核和你的 Agent 链路全部正常。

第 2 步:跑一次官方试玩,理解拍板机制
3

Step 3: Connect to the notification channel and let the cards come to you第 3 步:接上通知渠道,让卡片来找你

The decision to try it out is made on the terminal, but the real value lies in "not being on the terminal": Open the configuration file, direct the notification channel plugin to at least one of Feishu, DingTalk, WeChat Work, Slack, Discord, or system desktop notifications, and fill in the webhook according to the corresponding template in the "Plugin System" documentation of the repository.

Once the setup is complete, send any task that involves asking a question midway through the meeting for testing: you can close the terminal window directly after the task is sent. When the card appears on your phone, reply with the number. The kernel will automatically pick up the conversation and complete the rest of the task.

To see the overall progress, type agent-autoloop status at any time. Today's number of rounds and token expenditure are clearly visible, with no need to check any back-end interface – this project doesn't have a back-end interface at all.

中文

试玩的拍板在终端里完成,但真正的价值是「人不在终端」:打开配置文件,把通知通道插件指向飞书、钉钉、企业微信、Slack、Discord 或系统桌面通知中的至少一个,按仓库「插件体系」文档里对应模板填 webhook 即可。

配好后随便派一个会中途提问的任务测试:任务发出后你可以直接关掉终端窗口。等卡片出现在手机上,回复序号,内核会自动接上会话把剩下的活干完。

想看全局进展,任何时候敲 agent-autoloop status;今天的轮数、token 花费一目了然,不用翻任何后台界面——这个项目压根没有后台界面。

第 3 步:接上通知渠道,让卡片来找你
4

Step 4: Assign the first real task第 4 步:派第一个真实任务

Start with low-risk tasks, such as "Sort through yesterday's unread emails once every day at 8 a.m. and generate summary cards". Rewrite the task description into your own version based on the "ready-to-use scenario templates" in the repository, and submit it to the kernel queue.

Observe the first run: Event → Agent executes in the background → Card is approved → Task is completed. If a certain step is not going smoothly, use status to see where it is stuck. Usually, it is because a required field is missing from the corresponding plugin configuration.

Once the process is stable, increase the task density and integrate the "Code Deployment Pipeline" plugin according to the documentation to achieve a fully automated closed loop from requirement cards to merge requests.

中文

从低风险任务开始,比如「每天早八点整理一次昨日未读邮件并产出摘要卡片」。把任务描述按仓库里「拿来即用的场景模板」改写成你的版本,交给内核排队。

观察第一轮运行:事件 → Agent 后台执行 → 卡片拍板 → 干完收工。如果某个环节不顺,用 status 看卡在哪,一般是对应插件配置缺一项必填字段。

跑稳之后再加任务密度,并按文档把「代码上线流水线」插件接进来,实现从需求卡片到合并请求的全自动闭环。

第 4 步:派第一个真实任务
5

Problem Encountered遇到问题

The agent-autoloop command does not exist: Check whether Node 24 is installed and whether the installation script is executed with the current user's permissions. Cards cannot be received: Check the webhook and secret key of the notification plugin, and send a test message first. The task is stuck and does not progress: Check the cards awaiting approval in the status, and use the answer command to provide a response. The Agent Autoloop source code package that accompanies this article can be found via the download button above.

中文

agent-autoloop 命令不存在:检查 Node 24 是否装好、安装脚本是否以当前用户权限执行;卡片收不到:核对通知插件 webhook 与密钥,先发一条测试消息;任务卡住不推进:status 里看等待拍板的卡片,用 answer 命令补答即可。本文配套 Agent Autoloop 源码包见上方下载按钮。

资源下载 · Download

⬇ Download · 点击下载:Agent Autoloop v1.9.3 完整源码包(本教程配套)

MIT · GitHub 仓库 zip · 教程全程使用本包,无需另外下载

0阅读0 条评论

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。

评论 文明发言,让讨论更有价值

正能量公益广告今日正能量学一点,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