MCP Hub
A TypeScript gateway for Model Context Protocol servers that aggregates multiple downstream MCP servers, providing tool discovery, routing, smart search, and recommendations.
README
MCP Hub
MCP Hub is a TypeScript gateway for Model Context Protocol servers. It lets you register multiple downstream MCP servers and expose one hub surface for tool discovery, routing, smart search, and recommendations.
Status: experimental. The project is suitable for local development and portfolio review, but the public API should be treated as pre-1.0.
What It Does
- Aggregates tool metadata across registered MCP servers.
- Routes tool calls to the selected downstream server.
- Provides hub-native tools for catalog browsing, delegated calls, natural-language search, and recommendations.
- Supports stdio and HTTP/SSE downstream server entries.
- Keeps downstream credentials outside the repository through environment variables or per-server env files.
- Includes unit, integration, e2e, performance, and registry hardening tests.
MCP Hub does not bundle third-party MCP servers. All server examples are placeholders that you replace with your own commands, URLs, and credential storage.
Requirements
- Node.js 18 or newer
- npm 10 or newer
- One or more MCP servers that you can run locally or reach over HTTP/SSE
Install
git clone https://github.com/ftaricano/mcp-hub.git
cd mcp-hub
npm ci
npm run build
When an npm package is published, the intended package name is @ftaricano/mcp-hub.
Quickstart
Create a local hub config:
cp hub-config.example.json hub-config.json
Edit hub-config.json so each server entry points to a real MCP server in your environment. Keep hub-config.json local; it is ignored by Git because it may contain paths or environment references that are specific to your machine.
Minimal stdio server example:
{
"servers": [
{
"id": "docs-server",
"name": "Documentation MCP Server",
"command": "node",
"args": ["/absolute/path/to/docs-mcp/dist/index.js"],
"envFile": "/absolute/path/to/docs-mcp/.env",
"inheritEnv": ["DOCS_REGION"],
"protocol": "stdio",
"enabled": true,
"timeout": 30000,
"retries": 2,
"toolCallRetries": {
"enabled": true,
"maxAttempts": 2,
"retryableTools": ["search_docs"]
},
"tags": ["docs", "knowledge"]
}
],
"cache": {
"enabled": true,
"ttl": 300000
},
"security": {
"rate_limit": 100,
"validate_schemas": true
},
"logging": {
"level": "info",
"format": "json"
}
}
Run the hub:
HUB_CONFIG="$PWD/hub-config.json" npm start
MCP Client Config
Register the built hub in any MCP client that supports stdio servers:
{
"mcpServers": {
"hub": {
"command": "node",
"args": ["/absolute/path/to/mcp-hub/dist/index.js"],
"env": {
"HUB_CONFIG": "/absolute/path/to/mcp-hub/hub-config.json"
}
}
}
}
See LMStudio-MCP-Config.md for an LM Studio example.
Hub Tools
MCP Hub exposes four hub-native tools:
list-all-tools: returns the aggregated downstream tool catalog.call-tool: calls a named tool on a named downstream server.smart-search: ranks available tools for a natural-language query.get-recommendations: returns recommendation-oriented results based on available tool metadata.
Configuration
HUB_CONFIG can point to any JSON file matching the hub config shape. If it is not set, the hub tries common local paths for hub-config.json.
Server entries can use:
commandandargsfor stdio MCP servers,urlandprotocol: "http"for HTTP/SSE entries,envfor non-secret static values,envFilefor per-server secrets or local credentials,inheritEnvfor explicit parent environment pass-through,toolCallRetriesfor known-idempotent tools only.
Downstream stdio servers inherit only a small runtime-safe environment baseline by default. Add any required parent variables explicitly with inheritEnv.
Architecture
flowchart LR
Client["MCP client"] --> Hub["MCP Hub"]
Hub --> Registry["Server registry"]
Hub --> Search["Smart search and recommendations"]
Registry --> ServerA["Docs MCP server"]
Registry --> ServerB["Issue tracker MCP server"]
Registry --> ServerC["Any other MCP server"]
Key modules:
src/index.ts: MCP server entry point and hub-native tool handlers.src/registry/: downstream server lifecycle, connections, and tool discovery.src/intelligence/: search and recommendation scoring.src/cache/: result and metadata caching.src/utils/redaction.ts: argument summarization without logging raw sensitive values.
Development
npm run type-check
npm run lint
npm run format:check
npm test
npm run build
Useful focused suites:
npm run test:unit
npm run test:integration
npm run test:e2e
npm run test:performance
Packaging
Validate the npm package contents before publishing:
npm pack --dry-run --json
The package is intentionally restricted with the files field in package.json; local configs, logs, tests, coverage, testing-interface/, and generated artifacts should not ship in the npm tarball.
Security
Do not commit secrets or local operational state. In particular, keep these files and directories out of Git:
.env,.env.*except.env.examplehub-config.jsonand machine-specific config variantstokens/,auth-state/,config/auth-state/credentials.json,token.json, private keys, certificates, logs, and JSONL traces
Report vulnerabilities privately through GitHub Security Advisories. See SECURITY.md for the full policy.
Troubleshooting
If the hub starts but downstream tools do not appear:
- confirm
HUB_CONFIGpoints to the intended file, - verify each downstream server
command,args,url, andprotocol, - run the downstream MCP server directly before registering it behind the hub,
- check whether the server requires variables in
envFileorinheritEnv, - keep mutating tool calls out of retry lists unless they are explicitly idempotent.
If build or tests fail after dependency changes:
- run
npm cifrom a clean checkout, - run
npm run type-checkbefore debugging runtime behavior, - remove stale
dist/output withnpm run clean && npm run build.
Contributing
Issues and focused pull requests are welcome. Before opening a PR, run:
npm run type-check
npm run lint
npm run format:check
npm test
npm run build
Please avoid committing local credentials, generated logs, screenshots, or machine-specific hub configs.
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。