mcp-home-server
A dynamic MCP server that hot-reloads tools at runtime and provides URL tracking, onboarding, and scouting features for AI clients.
README
mcp-home-server
Single, stable Model Context Protocol (MCP) "door" server that runs on your home machine and exposes a dynamic set of tools to every connected AI client (desktop, phone, future agents). Tools can be added, edited, or removed at runtime — no restart needed.
Behind the door:
- Dynamic tool loader watches
src/tools/and hot-reloads each module via chokidar + the MCP SDK'slist_changednotification. - JSON manifest is regenerated automatically and served at
GET /manifest. - URLTracker ingests "AI usage" events, hashes them, and emits a
url_tracker:new_aievent the first time a URL is seen. - Onboarding agent listens to those events and writes per-AI onboarding
payloads (
manifests/onboarding/<hash>.json) so new clients can instantly configure themselves to use this MCP host. - Scouting worker periodically scans configured roots (
SCOUT_ROOTS) for scripts/modules that look like MCP tool candidates; proposals are written tomanifests/pending/for human review (no auto-promotion). - Dedupe worker compares registered tools and writes overlap reports to
docs/dedupe/to surface merges/refactors.
Quick start
git clone https://github.com/manniegreenspan/mcp-home-server.git
cd mcp-home-server
npm install
npm run dev
# Server binds 127.0.0.1 on the first free port in 7000-7999.
# Check console for the actual URL, e.g.: http://127.0.0.1:7432/mcp
Test it with the MCP Inspector:
npx @modelcontextprotocol/inspector@latest
# Paste the MCP endpoint URL into the inspector.
Endpoints
| Path | Method | Purpose |
|---|---|---|
/mcp |
POST/GET | MCP Streamable HTTP endpoint (JSON-RPC) |
/manifest |
GET | JSON manifest of currently registered tools |
/url-track |
POST | URLTracker ingestion: { "url": "...", "source": "..." } |
/health |
GET | Liveness + last chokidar refresh |
Adding a tool
Drop a file in src/tools/:
// src/tools/hello-world.ts
import { z } from "zod";
import type { ToolModule } from "../lib/types";
export default {
name: "hello_world",
description: "Says hello to a name",
inputSchema: { name: z.string().describe("Name to greet") },
handler: async ({ name }) => ({
content: [{ type: "text", text: `Hello, ${name}!` }]
})
} satisfies ToolModule;
The loader picks it up immediately, registers it with the MCP server, and
sends the notifications/tools/list_changed notification. Connected clients
(Inspector, Claude Desktop, etc.) refresh their tool list automatically.
Delete the file → tool unregistered. Edit the file → tool re-registered.
Cloudflare deployment
See docs/DEPLOYMENT.md. v1 recommends Cloudflare
Tunnel fronting the local Express server. A pure-Cloudflare-Worker rewrite
is in the ROADMAP (Workers can't do dynamic import() of src/tools/*.ts).
Project layout
mcp-home-server/
├── src/
│ ├── server.ts # main entry: Express + McpServer + transports
│ ├── ports.ts # port scan 7000-7999
│ ├── manifest.ts # generate/serve JSON manifest
│ ├── logger.ts # pino instance
│ ├── event-bus.ts # shared EventEmitter
│ ├── tools/ # dynamic tool modules (one file per tool)
│ ├── workers/ # onboarding-agent, scouting-worker, dedupe-worker
│ ├── url-tracker/ # ingestion + detector + hash store
│ └── lib/ # types, hashing
├── manifests/
│ ├── current.json # auto-regenerated manifest
│ ├── pending/ # scout proposals (manual review)
│ └── onboarding/ # onboarding payloads written by agent
├── docs/
│ ├── DEPLOYMENT.md
│ ├── ARCHITECTURE.md
│ └── dedupe/ # dedupe-worker reports
├── tests/
└── logs/
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。