hivemtk
开源私域 AI 营销 SCRM 系统,Martech 获客引流利器!全渠道聚合(抖音/小红书/快手/微信/WhatsApp/TikTok/Telegram)统一收件箱与 CDP 客户管理。🤖核心功能:1. ReAct 自主 AI 智能体(41原子工具/销冠 SOP 自动化编排);2. 本地知识库 RAG(混合检索/混合粗排);3. 独家 GEO 模块(生成式大模型引擎优化/AI 搜索占位/内容创作),让 ChatGPT/Perplexity/DeepSeek 优先推荐你的品牌!🔒数据安全:支持 100% 私域自托管部署,数据零出域
原始仓库简介 · 中文摘要待补充 · ai-agent / cdp / llm / marketing-technology / martech / open-source
README / 项目原文
🐝 HiveMtk · 开源私域营销 AI 系统
把销冠的能力,复制给团队里每一个普通人。
一个工作台管所有社媒,一个 AI 智能体谈所有客户,数据一寸不出门。
开源 SCRM · ReAct 自主智能体 · 10+ 渠道触达 · 数据零出域 · 5 分钟自托管部署
数字一览
| 94 个业务模块 SCRM+AI+CDP+自动化 |
10+ 触达渠道 抖音/小红书/TikTok/WhatsApp… |
45 个智能体工具 ReAct 自主编排 |
0 条数据出域 本地推理,零外网可跑 |
界面一览
| 统一收件箱 | AI 智能体 |
|---|---|
![]() |
![]() |
| 营销画布 | 客户 360° |
![]() |
![]() |
| 知识库 RAG | 数据大屏 |
![]() |
![]() |
这是什么
HiveMtk("Hive"+"Marketing Toolkit")是一套面向中文私域运营场景的开源 AI 营销系统,把四件事在同一个仓库里做透:
- 多端社媒触达 —— 抖音/快手/小红书/闲鱼/TikTok 经 Chrome 扩展桥接,企微/Telegram/WhatsApp/邮件/短信协议直连,统一收件箱
- ReAct 自主智能体 —— 感知 → 规划 → 调工具 → 反思,45 个原子工具自主组合,不是写死的 if-else 工作流
- 本地知识库 RAG —— pgvector 1024 维混合检索 + bge-m3 + bge-reranker-v2-m3 精排
- 数据零出域 —— llama.cpp(Qwen2.5)+ TEI 本地推理栈,对话/知识库/向量化全程在客户内网完成
覆盖获客 → 触达 → 转化 → 复购全链路。不是"给大模型套个壳",是直接能跑营销业务的完整系统;另附 GEO 模块打通 AI 搜索获客闭环。完整功能列表与已知限制如实披露,见 docs/marketing-features/README.md。
适合谁:5–50 人的成长型团队 · 合规敏感行业(金融/医疗/政企)· 不想被 SaaS 厂商锁定的运营团队。
数字背后的差异
诚实对比,敢暴露劣势。没有银弹,按需选择。
| 维度 | HiveMtk(本项目) | Dify / FastGPT | 商业 SCRM(微伴/尘锋) | 源雀 / MoChat |
|---|---|---|---|---|
| 核心定位 | 私域营销 AI 系统 | 通用 LLM 应用平台 | 商业 SaaS | 企微 SCRM |
| 触达端 | 多端(10+) | 无内置 | 1-3 端 | 1 端(企微) |
| AI 能力 | ReAct 智能体 + 45 工具 | 可视化 Workflow | 基础客服机器人 | 简单 RAG / 无 |
| 数据部署 | 100% 私域 + 本地推理栈 | 自托管 / SaaS | 云端 | SaaS / 私有 |
| 开源协议 | AGPL-3.0 | Apache-2.0 | 闭源 | 部分开源 |
| 适合谁 | 想自主可控、多端、强 AI 的团队 | 纯 AI 应用开发 | 无 IT 团队的中小商家 | 企微深度用户 |
核心能力
🌐 多端触达:一个工作台全管
| 渠道 | 触达 | 智能卡片 | 自动回复 | RAG 客服 | 备注 |
|---|---|---|---|---|---|
| 抖音 / 快手 / 小红书 / 闲鱼 / TikTok | ✅ | ✅ | ✅ | ✅ | Chrome 扩展桥接,你自己的登录态浏览器,无需无头浏览器 |
| 微信 / 企业微信 | ✅ | — | ✅ | ✅ | 含社群/朋友圈 |
| Telegram | ✅ | — | ✅ | ✅ | Bot 协议直连,群线索挖掘 + 主动 DM + 入群管控 |
| ✅ | — | ✅ | ✅ | Cloud API + 模板消息 | |
| 邮件 / 短信 | ✅ | — | ✅ | — | SMTP;阿里云/腾讯云/华为云 |
统一 CDP 客户视图,一份资料全渠道触达;统一收件箱,会话/工单/留言一处看完。渠道能力明细与 Telegram 五大获客能力,见 docs/marketing-features/README.md。
🤖 ReAct 自主智能体:不是写死的工作流
- ReAct 循环:感知 → 规划 → 调工具 → 反思,撞到预设之外的场景也能自己组合工具搞定
- 45 个原子工具:触达/项目管理/客户/知识库/业务五域,Permission→Retry→Timeout→RateLimit→Audit 五层装饰器逐层包裹
- 混合检索 RAG:pgvector HNSW + BM25 RRF 融合粗排,bge-reranker 精排,可选 HyDE/MultiQuery 改写
- 多智能体协作:被动应答智能体 + 主动触达智能体;四层记忆系统(短期/长期/SOP 状态/业务记忆)
- 可视化工作流:营销自动化编辑器,零代码搭建销冠 SOP
工具注册逻辑以源码为准:user-server/internal/aiagent/agent/tooluse/(代码即清单)。
🔒 数据安全:100% 私域,对话不出客户内网
- 本地推理栈:llama.cpp(Qwen2.5)+ TEI(bge-m3 + bge-reranker-v2-m3),LLM/Embedding/Rerank 三个 OpenAI 兼容服务跑在客户内网
- 零外网可跑:对话、知识库、向量化、检索增强全程域内完成
- FRP 私域穿透:访客从公网进,数据经隧道回本地,云端不落一条对话
- 合规友好:满足等保、数据出境管控、私有化部署基线
- 可选云端 LLM:把
LLM_BASE_URL指向 DeepSeek/OpenAI 即可,Embedding/Rerank 仍强制本地
🎯 GEO 智能优化:AI 搜索获客闭环
SEO 争搜索结果排名,GEO(生成式引擎优化)争 AI 答案里的"席位"——让 ChatGPT Search、Perplexity、Google SGE 作答时提及你的品牌。
品牌配置 → 关键词蒸馏 → 内容创作 → 多模型验证 → 平台同步发布
↑ ↓
数据增强 ← 历史验证数据 ←──────────┘
关键词蒸馏 · E-E-A-T + Schema 注入 · 多厂商 LLM 模拟 AI 搜索作答量化"席位" · RAG 锚定品牌事实 · 高权重平台同步。完整指南见 user-server/docs/geo-module-guide.md。
适用场景
| 场景 | 关键能力 | 典型行业 |
|---|---|---|
| 私域获客 → AI 自动谈单 | 多端线索接入 + ReAct 智能体承接咨询 + 异议处理 + 逼单邀约 | 招商加盟 / 医美 / 教培 / B2B |
| 企微 SCRM 多账号聚合 | 多账号统一收件箱 + 客户资产沉淀 + 离职继承 | 连锁门店 / 品牌方 |
| AI 客服 / RAG 知识库 | 混合检索 + 7×24 自动回复 | 电商 / SaaS / 售后 |
| 营销自动化 SOP | 销冠 SOP 可视化编排 + RFM 分层 + 流失预警 + 沉睡激活 | 会员制零售 / 母婴 / 美业 |
| 跨境出海多渠道触达 | TikTok + WhatsApp + 邮件 + Telegram 统一管理 | 跨境电商 / 出海品牌 |
| 合规私有化部署 | 本地推理栈 + 零出域 + 行级权限 + 审计存档 | 金融 / 医疗 / 政企 |
🚀 5 分钟跑起来
唯一要求:Docker 24+ 与 Docker Compose v2。硬件最低 4 核/8GB/50GB(dev 档),推荐 8 核/16GB(prod 档含 14B 模型);GPU 可选,无 GPU 走 CPU 推理。
# 1. 克隆(国内走 Gitee,海外走 GitHub 镜像)
git clone https://gitee.com/xhpmayun/hivemtk.git && cd hivemtk
# 2. 一键安装:.env 模板 + 构建前后端 + 下载模型 + 拉起数据层与推理栈
make install
# 3. 改密钥
vim .env # 必改:POSTGRES_PASSWORD / REDIS_PASSWORD / JWT_SECRET
# 只有把平台端跑起来(PLATFORM_ENABLED=true)才用到:PLATFORM_ADMIN_PASSWORD / MERCHANT_API_SECRET
# 4. 启动
make dev # user-server 热更新 → http://localhost:8204
cd user-web && npm run dev # 前端工作台 → http://localhost:8211(另开终端)
dev/prod 模型档切换、常见命令、数据库备份恢复等运维细节,见 docs/DEPLOYMENT_GUIDE.md 与 docs/TROUBLESHOOTING.md。
FAQ
和 Dify / FastGPT 有什么区别? HiveMtk 是私域营销系统,聚焦"多端触达 + 销冠 SOP + CDP + 零出域",内置 94 个业务模块开箱即用;Dify/FastGPT 是通用 LLM 应用开发平台,偏 Workflow 编排,不内置营销业务。
是真的开源吗?可以商用吗? ✅ AGPL-3.0 完全开源。可自由 fork、私有部署、二次开发、商用。唯一限制:修改后的版本若通过网络对外提供服务(SaaS/云端 API),必须同样按 AGPL-3.0 开源你的修改;仅内部私有部署无需公开。
数据真的不出域吗? ✅ 所有对话、知识库、向量化、检索增强全程在客户内网完成。本地推理栈跑在客户内网,FRP 穿透时云端不落一条对话。即使配置云端 LLM,也只传 prompt 文本,PII 本地脱敏后再调用。
支持哪些大模型? ✅ DeepSeek / 通义千问 / GPT-4o / 智谱 GLM / 本地 Qwen2.5。LLM 路由网关按场景动态路由:复杂异议用强模型,常规回复用轻模型。Embedding/Rerank 强制本地。
海外能用吗? ✅ TikTok + WhatsApp + Telegram + Email 适配跨境出海场景。
有 SaaS 版吗?
❌ 不提供。坚持私有化部署、数据自主可控。企业级技术支持/定制集成联系 jideilvluoqun@gmail.com。
架构一图
访客浏览器(公网)
│ HTTPS / WSS(经 FRP / 公网 IP / 反代)
▼
┌─────────────────────────────────────────────────┐
│ 客户本地内网 │
│ │
│ user-server (Go + Gin) :8204 │
│ ├── PostgreSQL + pgvector :8202 (1024 维) │
│ └── Redis 7 :8203 │
│ │
│ 宿主机推理栈(llama.cpp,非容器化) │
│ ├── LLM :8207 (Qwen2.5) │
│ ├── Embedding :8208 (bge-m3) │
│ └── Rerank :8209 (bge-reranker-v2-m3) │
│ │
│ user-web (Vue 3) · embed-sdk(客服 Widget) │
└─────────────────────────────────────────────────┘
┆ HTTPS(低频心跳:install_id / 版本号 / 主机信息 / 运行指标 / 设备指纹)
┆ —— 这条边默认不存在:平台端是可选本地组件,
┆ 未设 PLATFORM_ENABLED=true 时一个请求都不发
┆
▼
平台端(独立仓库 hivemtk-platform,可选自建):仅元数据,
不接触、不存储、不访问任何业务数据;无任何商户授权/标识校验环节
后端严格遵循五层架构(Controller → Service → Repository → Model → DTO,禁止跨层调用,CI 强制检查)。完整规范见 docs/architecture/GO_FIVE_LAYER_ARCHITECTURE.md,与平台端的分工见 docs/architecture/部署方案_用户端.md。
技术栈:Go 1.26 + Gin + GORM · Vue 3 + Vite + Element Plus + Pinia · PostgreSQL 15 + pgvector · Redis 7 · llama.cpp + Qwen2.5 · TEI + bge-m3 / bge-reranker-v2-m3 · Docker Compose。
仓库结构
hivemtk/ # 用户端仓库
├── user-server/ # Go 后端(核心业务,五层架构)→ user-server/README.md
├── user-web/ # Vue 3 前端(B 端工作台)→ user-web/README.md
├── embed-sdk/ # 嵌入式客服 Web Widget(IIFE/ESM)
├── website/ # 官网源码(Vue 3 SPA)→ 由 .github/workflows/website-pages.yml 发到 GitHub Pages
├── migrations/ # 数据库迁移 SQL(幂等)
├── scripts/inference-host/ # 宿主机推理栈脚本(llama.cpp + TEI)
├── docs/ # 文档:INDEX.md / DEPLOYMENT_GUIDE.md / marketing-features/ …
├── docker-compose.yml # 数据层编排(PG + Redis)
├── Makefile # 一键安装/启动/停止/推理栈/开发热更新
└── .env-example # 环境变量模板
文档导航
| 文档 | 入口 |
|---|---|
| 📖 英文说明 | README.en.md |
| 📚 文档索引 | docs/INDEX.md |
| 🚀 部署指南 / 排查手册 | docs/DEPLOYMENT_GUIDE.md · docs/TROUBLESHOOTING.md |
| 📦 94 个功能模块明细 | docs/marketing-features/README.md |
| 📕 功能文档 · 操作手册(人读 / AI 读 四册) | docs/handbook/README.md |
| 🧭 路线图 | .github/ROADMAP.md |
| 🤝 贡献指南 / 安全策略 | CONTRIBUTING.md · SECURITY.md |
合规与免责
主动触达(短信/邮件/社媒私信/Telegram/WhatsApp 等主动推送能力)属于核心敏感功能,本项目作为开源工具不对使用者的调用方式承担责任;每次主动触达发送,服务端日志都会打印不可关闭的 [COMPLIANCE] 合规提示。完整声明见 DISCLAIMER.md(English)。
License:本项目采用 AGPL-3.0 发布。私有部署、内部使用、二次开发完全自由、无需开源;修改代码并以 SaaS/云服务/托管 API 对外提供时,必须按 AGPL-3.0 开源修改部分。版权与联系方式见 NOTICE。
联系与社区
| 渠道 | 入口 | 说明 |
|---|---|---|
| 🐛 Bug / Feature Request | Gitee Issues / GitHub Issues | 提交问题与建议 |
| 💬 微信交流群 | 通过 Issue / 邮箱申请加入 | 产品/技术/运营答疑,禁止广告/政治/人肉,违者秒踢 |
| 📧 商务合作 / 技术支持 | jideilvluoqun@gmail.com |
企业级技术支持、定制集成、私有部署咨询 |
| 🔒 安全漏洞 | jideilvluoqun@gmail.com |
私密报告,详见 SECURITY.md |
添加作者微信,入群同频交流
扫码或搜索微信号
xiao142000,备注来意通过,一起交流。

镜像仓库:Gitee 主仓库 gitee.com/xhpmayun/hivemtk(国内推荐,下载更快)· GitHub 镜像 github.com/xiaofang142/hivemtk(Actions 定时同步)
⭐ Star / Watch 一下,跟项目一起成长
多端打透 · AI 真自主 · 数据封死在域内
Made with ❤️ by HiveMtk Team
🐝 HiveMtk · Open-Source Self-Hosted AI Marketing System
Turn your best salesperson's playbook into every rep's daily workflow.
One workspace for every social channel. One AI agent for every customer conversation. Zero data leaving your perimeter.
Open-Source SCRM · ReAct Autonomous Agent · 10+ Channels · Zero Data Egress · 5-Minute Self-Hosting
At a Glance
| 94 business modules SCRM + AI + CDP + automation |
10+ reach channels Douyin / Xiaohongshu / TikTok / WhatsApp… |
45 agent tools ReAct autonomous orchestration |
0 bytes of data egress Local inference, fully offline-capable |
Screenshots
| Unified Inbox | AI Agent |
|---|---|
![]() |
![]() |
| Marketing Canvas | Customer 360° |
![]() |
![]() |
| Knowledge Base RAG | Dashboard Screen |
![]() |
![]() |
What Is This
HiveMtk ("Hive" + "Marketing Toolkit") is an open-source AI marketing system for private-domain operations, nailing four things in a single repo:
- All-channel social reach — Douyin / Kuaishou / Xiaohongshu / Xianyu / TikTok via a Chrome-extension bridge; WeCom / Telegram / WhatsApp / Email / SMS via protocol-direct; one unified inbox
- ReAct autonomous agents — perceive → plan → tool-call → reflect; 45 atomic tools composed on the fly, not hardcoded if-else workflows
- Local knowledge base RAG — pgvector 1024-dim hybrid retrieval + bge-m3 + bge-reranker-v2-m3 fine ranking
- Zero data egress — llama.cpp (Qwen2.5) + TEI local inference stack; conversations, knowledge base, and embeddings never leave your network
Covers the full acquisition → outreach → conversion → repurchase funnel. Not an LLM wrapper — a complete system that runs real marketing operations, plus a GEO module closing the AI-search acquisition loop. The full feature list (and known limitations) is honestly disclosed in docs/marketing-features/README.md.
Who it's for: growing teams of 5–50 · compliance-sensitive industries (finance / healthcare / government) · operators done with SaaS vendor lock-in.
How It Compares
Honest comparison, weaknesses included. No silver bullets — pick what fits.
| Dimension | HiveMtk (this repo) | Dify / FastGPT | Commercial SCRM | SourceQue / MoChat |
|---|---|---|---|---|
| Core positioning | AI marketing system | General LLM app platform | Commercial SaaS | WeCom SCRM |
| Reach channels | Multi-channel (10+) | None built-in | 1-3 | 1 (WeCom) |
| AI capability | ReAct agents + 45 tools | Visual Workflow | Basic CS bot | Simple RAG / none |
| Data deployment | 100% on-prem + local inference | Self-host / SaaS | Cloud | SaaS / private |
| License | AGPL-3.0 | Apache-2.0 | Proprietary | Partially open |
| Best for | Teams wanting self-control, multi-channel, strong AI | Pure AI app dev | SMBs without IT | WeCom power users |
Core Capabilities
🌐 All-Channel Reach: One Workspace
| Channel | Outreach | Smart Cards | Auto-Reply | RAG CS | Notes |
|---|---|---|---|---|---|
| Douyin / Kuaishou / Xiaohongshu / Xianyu / TikTok | ✅ | ✅ | ✅ | ✅ | Chrome-extension bridge over your own logged-in browser — no headless browser |
| WeChat / WeCom | ✅ | — | ✅ | ✅ | Groups + Moments |
| Telegram | ✅ | — | ✅ | ✅ | Bot protocol direct: group-lead mining + proactive DM + gate control |
| ✅ | — | ✅ | ✅ | Cloud API + template messages | |
| Email / SMS | ✅ | — | ✅ | — | SMTP; Aliyun / Tencent / Huawei |
Unified CDP customer profiles, one profile reaching everywhere; unified inbox — every conversation, ticket, and DM in one place. Per-channel details and Telegram's five acquisition capabilities: docs/marketing-features/README.md.
🤖 ReAct Autonomous Agents: Not Dead Workflows
- ReAct loop: perceive → plan → tool-call → reflect; unseen scenarios get handled by composing tools on the fly
- 45 atomic tools across five domains (reach / projects / customers / knowledge base / business), each wrapped by a five-layer decorator chain: Permission → Retry → Timeout → RateLimit → Audit
- Hybrid RAG: pgvector HNSW + BM25 with RRF fusion, bge-reranker fine ranking, optional HyDE / MultiQuery rewrite
- Multi-agent collaboration: reactive answering agent + proactive outreach agent; four-layer memory (short-term / long-term / SOP state / business facts)
- Visual workflow builder: zero-code SOP editor for marketing automation
Tool registration logic, per source: user-server/internal/aiagent/agent/tooluse/ (code as inventory).
🔒 Data Security: 100% On-Prem, Zero Egress
- Local inference stack: llama.cpp (Qwen2.5) + TEI (bge-m3 + bge-reranker-v2-m3) — three OpenAI-compatible services (LLM / Embedding / Rerank) inside your network
- Fully offline-capable: conversations, knowledge base, embeddings, and RAG all complete within your perimeter
- FRP private tunneling: visitors come in from the public internet, data flows back through the tunnel — the cloud never stores a single message
- Compliance-friendly: meets data-residency and private-deployment baselines
- Optional cloud LLM: point
LLM_BASE_URLat DeepSeek/OpenAI for stronger models; Embedding/Rerank stay strictly local
🎯 GEO Optimization: AI-Search Acquisition Loop
SEO fights for search ranking; GEO (Generative Engine Optimization) fights for a seat in the AI answer — getting your brand cited when ChatGPT Search, Perplexity, or Google SGE answers.
Brand config → Keyword distillation → Content creation → Multi-model validation → Platform syndication
↑ ↓
Data feedback ← Historical validation data ←──────────────┘
Keyword distillation · E-E-A-T + Schema injection · multi-vendor LLM simulation quantifying your "AI answer seat" · RAG grounding to brand facts · high-authority platform syndication. Full guide: user-server/docs/geo-module-guide.md.
Typical Use Cases
| Scenario | Key Capabilities | Industries |
|---|---|---|
| Private-domain acquisition → AI auto-selling | Multi-channel lead intake + ReAct agent handles inbound + objection handling + closing | Franchise / Medical aesthetics / EdTech / B2B |
| WeCom SCRM multi-account aggregation | Unified inbox + customer asset sink + departure inheritance | Chain stores / Brands |
| AI customer service / RAG knowledge base | Hybrid retrieval + 7×24 auto-reply | E-commerce / SaaS / After-sales |
| Marketing automation SOP | Visual SOP editor + RFM segmentation + churn prediction + reactivation | Membership retail / Beauty |
| Cross-border multi-channel outreach | TikTok + WhatsApp + Email + Telegram unified | Cross-border e-commerce |
| Compliant on-premise deployment | Local inference + zero egress + row-level security + audit archive | Finance / Healthcare / Government |
🚀 Up and Running in 5 Minutes
Requirements: Docker 24+ with Docker Compose v2. Minimum hardware 4 cores / 8GB RAM / 50GB disk (dev tier); recommended 8 cores / 16GB (prod tier with 14B model). GPU optional — CPU-only inference works.
# 1. Clone (Gitee primary for China; GitHub mirror elsewhere)
git clone https://gitee.com/xhpmayun/hivemtk.git && cd hivemtk
# 2. One-click install: .env template + build frontends + download models + start data layer & inference stack
make install
# 3. Set 4 secrets
vim .env # POSTGRES_PASSWORD / REDIS_PASSWORD / JWT_SECRET / PLATFORM_ADMIN_PASSWORD
# 4. Start
make dev # user-server with hot reload → http://localhost:8204
cd user-web && npm run dev # Admin workspace → http://localhost:8211 (another terminal)
For dev/prod model tiers, common commands, and backup/restore operations, see docs/DEPLOYMENT_GUIDE.md and docs/TROUBLESHOOTING.md.
FAQ
How is it different from Dify / FastGPT? HiveMtk is a marketing system focused on "all-channel reach + sales-copilot SOP + CDP + zero egress", shipping 94 business modules out of the box. Dify / FastGPT are general LLM application platforms oriented around Workflow orchestration, without built-in marketing business.
Is it really open-source? Can I use it commercially? ✅ Fully open-source under AGPL-3.0. Fork, self-host, extend, and use commercially — free. The only constraint: if you offer a modified version as a network service (SaaS / cloud API), you must release your modifications under AGPL-3.0 as well. Internal-only self-hosting never triggers this.
Does data really stay on-premise? ✅ All conversations, knowledge base, embeddings, and RAG stay inside your network. The local inference stack runs on-prem; with FRP tunneling the cloud never stores a message. Even with a cloud LLM, only prompt text is sent, and customer PII is desensitized locally first.
Which LLMs are supported? ✅ DeepSeek / Qwen / GPT-4o / GLM / local Qwen2.5. The LLM routing gateway dispatches by scenario: strong models for complex objections, light models for routine replies. Embedding/Rerank are strictly local.
Can it be used overseas? ✅ TikTok + WhatsApp + Telegram + Email cover cross-border scenarios.
Is there a SaaS version?
❌ No. We insist on private deployment and data sovereignty. For enterprise support or custom integration: jideilvluoqun@gmail.com.
Architecture at a Glance
Visitor Browser (Public Internet)
│ HTTPS / WSS (via FRP / public IP / reverse proxy)
▼
┌─────────────────────────────────────────────────┐
│ Customer On-Prem Network │
│ │
│ user-server (Go + Gin) :8204 │
│ ├── PostgreSQL + pgvector :8202 (1024-dim) │
│ └── Redis 7 :8203 │
│ │
│ Host Inference Stack (llama.cpp, no containers)│
│ ├── LLM :8207 (Qwen2.5) │
│ ├── Embedding :8208 (bge-m3) │
│ └── Rerank :8209 (bge-reranker-v2-m3) │
│ │
│ user-web (Vue 3) · embed-sdk (CS Web Widget) │
└─────────────────────────────────────────────────┘
┆ HTTPS (low-frequency heartbeat: install_id / version / host info /
┆ runtime metrics / device fingerprint)
┆ -- this edge does not exist by default: the platform is an
┆ optional local component; with PLATFORM_ENABLED unset
┆ not a single request leaves the box
▼
Platform (optional, self-hosted from the separate hivemtk-platform repo):
metadata only. Never touches, stores, or accesses your business data.
There is no merchant licensing and no merchant-identity check anywhere.
The backend strictly follows a five-layer architecture (Controller → Service → Repository → Model → DTO, no cross-layer calls, CI-enforced). Full spec: docs/architecture/GO_FIVE_LAYER_ARCHITECTURE.md; deployment split with the platform: docs/operations/MERCHANT_DEPLOYMENT.md.
Tech stack: Go 1.25 + Gin + GORM · Vue 3 + Vite + Element Plus + Pinia · PostgreSQL 15 + pgvector · Redis 7 · llama.cpp + Qwen2.5 · TEI + bge-m3 / bge-reranker-v2-m3 · Docker Compose.
Repository Layout
hivemtk/ # User-side repo
├── user-server/ # Go backend (core business, 5-layer arch) → user-server/README.md
├── user-web/ # Vue 3 frontend (B-side workspace) → user-web/README.md
├── embed-sdk/ # Embeddable CS Web Widget (IIFE/ESM)
├── migrations/ # DB migration SQL (idempotent)
├── scripts/inference-host/ # Host inference stack scripts (llama.cpp + TEI)
├── docs/ # Docs: INDEX.md / DEPLOYMENT_GUIDE.md / marketing-features/ …
├── docker-compose.yml # Data layer orchestration (PG + Redis)
├── Makefile # install / up / down / inference / dev
└── .env-example # Env template
Documentation
| Doc | Entry |
|---|---|
| 📖 Chinese readme | README.md |
| 📚 Doc index | docs/INDEX.md |
| 🚀 Deployment guide / Troubleshooting | docs/DEPLOYMENT_GUIDE.md · docs/TROUBLESHOOTING.md |
| 📦 94 feature modules in detail | docs/marketing-features/README.md |
| 🧭 Roadmap | .github/ROADMAP.md |
| 🤝 Contributing / Security | CONTRIBUTING.md · SECURITY.md |
Compliance & Disclaimer
Proactive outreach (SMS / email / social DM / Telegram / WhatsApp) is a core sensitive capability; as an open-source tool, this project takes no responsibility for how users invoke it. Every proactive send prints an unclosable [COMPLIANCE] notice in server logs. Full statement: DISCLAIMER.md (English).
License: released under AGPL-3.0. Private deployment, internal use, and modification are fully free with no obligation to open-source; offering a modified version as a SaaS / cloud service / managed API triggers the copyleft obligation for your modifications. See NOTICE for copyright and contact.
Contact & Community
| Channel | Entry | Notes |
|---|---|---|
| 🐛 Bug / Feature Request | GitHub Issues / Gitee Issues | Report issues & suggestions |
| 💬 WeChat group | Apply via Issue / email | Product / tech / ops Q&A; no ads, no politics |
| 📧 Business / Support | jideilvluoqun@gmail.com |
Enterprise support, custom integration, deployment consulting |
| 🔒 Security reports | jideilvluoqun@gmail.com |
Private disclosure, see SECURITY.md |
Mirrors: Gitee primary gitee.com/xhpmayun/hivemtk (faster in China) · GitHub mirror github.com/xiaofang142/hivemtk (Actions-synced)
⭐ Star / Watch to follow along
All Channels · True AI Autonomy · Zero Data Egress
Made with ❤️ by HiveMtk Team





