完整的基于Docker的AI运维基础设施：Paperclip、n8n、Supabase、OpenClaw、ruflo和MCP
概览
这个仓库是架构和操作方法论：Docker 服务、操作手册，以及——只有在生产环境中运行数月代理才能学会的部分——如何保持一支代理舰队的存续。 代理存在于你服务器上的 tmux 会话中，你通过 Telegram 与它通信。它负责部署、抓取、编写工作流程、接听客户并运行设备。这里的所有内容都是源自实际运行操作的部分。 n8n/workflow/导出数据是从生产环境中采集的，并由人工编辑。这些编辑会通过自动检查验证，因此请将导出数据视为可读取的样本，而非被清除的夹具以重复使用——也不要假设自己的导出数据也是如此。如果你贡献工作流程JSON，手写合成样本数据而不是发送捕获的流量：姓名、地址、银行信息、列表ID和跟踪URL都能通过简单的查找替换而存活下来，编码令牌则隐形保存。
概述（AI 翻译）
用一个VPS的聊天应用运营一个小型代理机构。 这个仓库是堆栈和运行的 方法论：Docker 服务、操作手册，以及——只有运行代理才能学会的部分 制作数月——如何维持一支舰队的生命。
代理存在于你服务器的 tmux 会话中，你通过 Telegram 与它对话。它部署，
抓取、编写工作流程、回答客户并运行设备。这里的所有内容都是从衍生出来的
到底是什么在操控整个行动。
在已发货的示例中。
n8n/workflows/出口是从生产中捕获的，且是 手工涂黑。这些涂黑没有任何自动检查验证，所以导出时要处理 作为样本来阅读，而不是被擦洗过的灯具重复使用——也不要对自己的设备抱持同样的假设 导出。如果你贡献工作流程JSON，请手写合成样本数据，而不是直接发送 捕获的流量：姓名、地址、银行信息、列表ID和追踪URL均存活下来 天真的“查找并替换”，编码的代币则能隐形地存活下来。
| 🐣 从没搭建过服务器？ | **文档/从零开始设置。 |
本页由 AI 翻译，技术细节请以原文为准。
README
以下为原文 README 节选，完整内容请查看原文。
Stratoma AI Stack
Run a small agency from a chat app, on one VPS. This repo is the stack and the operating methodology: the Docker services, the playbooks, and — the part you only learn by running agents in production for months — how to keep a fleet of them alive.
The agent lives in a tmux session on your server and you talk to it from Telegram. It deploys,
scrapes, writes workflows, answers customers, and runs the box. Everything here is derived from
what actually runs the operation.
On the shipped examples. The
n8n/workflows/exports are captured from production and were redacted by hand. That redaction is not verified by any automated check, so treat the exports as samples to read, not as scrubbed fixtures to reuse — and do not assume the same of your own exports. If you contribute workflow JSON, hand-write synthetic sample data rather than shipping captured traffic: names, addresses, bank details, listing IDs and tracking URLs all survive a naive find-and-replace, and encoded tokens survive it invisibly.
| 🐣 Never set up a server before? | docs/SETUP-FROM-SCRATCH.md — every command, copy-paste, from buying a server to texting an AI agent that runs it for you. |
| 🐳 You know Docker? | Quick Start below. |
Hosting is one small VPS (~€20/mo at the time of writing), plus the Claude subscription the agent
runs on. There is no ANTHROPIC_API_KEY anywhere in this repo — see
Cost & security.
🎮 Showcase — built on this exact stack
Pokémon Madrid is a full Pokémon-style game set in Madrid, built almost entirely by AI from a prompt (Claude Code · Opus + Gemini for the art), and deployed on this stack: Hetzner + Coolify + self-hosted Supabase (account login & cloud saves). The whole thing — code, art, audio, deploys, testing — was directed from Telegram, with a group of friends sending photos and audio as context.
→ Get €20 free credit on Hetzner and build your own.
Documentation
| Doc | What it covers |
|---|---|
| SETUP-FROM-SCRATCH | Zero to a server you can text. Buy the VPS, SSH in, install Claude Code, tmux, the official Telegram plugin, a BotFather bot, then MCP tools, skills and DNS/API tokens. |
| CLAUDE-CODE-TELEGRAM-WORKFLOW | Day-to-day operating: the exact start command, tmux + systemd, the pairing flow for teammates, the four memory layers, the MCP catalog, security posture, troubleshooting. |
| MULTI-CLAUDE-MOTHER-AND-CHILDREN | Many projects, one panel: per-UNIX-user isolation, cross-tmux orchestration, the Supabase journal schema, call-transcript ingestion, adding a new project. |
安装
npx -y n8n-mcp配置
{
"mcpServers": {
"n8n-mcp": {
"type": "stdio",
"command": "npx",
"args": ["-y", "n8n-mcp"],
"env": {
"MCP_MODE": "stdio",
"LOG_LEVEL": "error",
"DISABLE_CONSOLE_OUTPUT": "true",
"N8N_API_URL": "${N8N_URL}",
"N8N_API_KEY": "${N8N_API_KEY}",
"N8N_MCP_TELEMETRY_DISABLED": "true"
}
},
"github": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}" }
}
}
}