mcp-hub
A lightweight control plane and local stdio agent that enables multiple MCP clients to connect to one local agent while configuration is managed from a remote control plane, supporting tool naming, exposure control, and multiple upstream source types.
README
mcp-hub
mcp-hub is a lightweight control plane and local stdio agent for managing MCP sources across AI coding tools.
It lets Codex, Claude Code, OpenCode, and other MCP clients connect to one local agent while the actual upstream MCP services are configured from a remote control plane.
- npm package:
@a1ua/mcp-hub - Repository: https://github.com/aiua-dev/mcp-hub
- License: MIT
Overview
MCP tools are often installed and configured per client. As the number of tools grows, keeping every local client in sync becomes repetitive and error-prone.
mcp-hub separates the system into two parts:
- Local agent
- Runs as a
stdioMCP server. - Loads workspace configuration from a control-plane API.
- Connects to upstream MCP services from the local machine.
- Falls back to cached configuration when the control plane is temporarily unavailable.
- Runs as a
- Control plane
- Maintains MCP sources, exposed tools, and hosted source definitions.
- Provides configuration endpoints for local agents.
- Includes a web console for managing sources and copied client configs.
Features
- One local
stdioentrypoint for multiple MCP clients. - Remote workspace configuration with local cache fallback.
- Tool naming and exposure control.
- Support for multiple upstream source types:
direct-httplocal-stdiohosted-npmhosted-single-file
- Codex-friendly TOML output with explicit
type = "stdio". - Shell-wrapped
npxcommand generation for GUI clients with limited$PATH. - Optional export profiles for exposing different source sets as separate MCP servers.
Install
Use the published local agent directly with npx:
npx -y @a1ua/mcp-hub \
--base-url https://your-control-plane.example.com \
--workspace mcp-hub \
--token your-token
Or install it globally:
npm install -g @a1ua/mcp-hub
Client Setup
Codex
codex mcp add mcp-hub -- \
npx -y @a1ua/mcp-hub \
--base-url https://your-control-plane.example.com \
--workspace mcp-hub \
--token your-token
You can also write the server configuration directly:
[mcp_servers."mcp-hub"]
type = "stdio"
command = "/bin/sh"
args = ["-lc", "PATH=\"/opt/homebrew/bin:/usr/local/bin:$PATH\"; if [ -d \"$HOME/.nvm/versions/node\" ]; then for dir in \"$HOME\"/.nvm/versions/node/*/bin; do [ -d \"$dir\" ] && PATH=\"$dir:$PATH\"; done; fi; cd \"$HOME\"; exec 'npx' '-y' '@a1ua/mcp-hub' '--base-url' 'https://your-control-plane.example.com' '--workspace' 'mcp-hub' '--token' 'your-token'"]
Config URL Mode
Use --config-url when the control plane gives you a full workspace or export config endpoint:
npx -y @a1ua/mcp-hub \
--config-url https://your-control-plane.example.com/v1/workspaces/mcp-hub/config \
--workspace mcp-hub \
--token your-token
CLI Options
| Option | Description |
|---|---|
--base-url |
Control-plane base URL. The agent resolves /v1/workspaces/<workspace>/config. |
--config-url |
Full configuration URL. Takes precedence over --base-url. |
--workspace |
Workspace ID and local cache key. |
--token |
Bearer token passed directly on the command line. |
--token-env |
Environment variable name used to read the Bearer token. |
--cache-dir |
Custom local cache directory. |
The same settings can also be supplied through environment variables:
MCP_AGENT_BASE_URLMCP_AGENT_CONFIG_URLMCP_AGENT_WORKSPACEMCP_AGENT_TOKENMCP_AGENT_TOKEN_ENVMCP_AGENT_CACHE_DIR
Project Structure
.
├── apps
│ ├── control-plane-api # Fastify API and hosted-source runtime management
│ └── control-plane-web # React control-plane console
├── packages
│ ├── agent # Published local stdio agent
│ ├── runtime # MCP aggregation runtime
│ └── shared # Shared schemas and config mapping
└── docs # Architecture and product notes
Development
Install dependencies:
pnpm install
Run the development helper:
pnpm go
Run API and web console separately:
pnpm dev:api
pnpm dev:web
Run checks:
pnpm -r test
pnpm build
Control-Plane Contract
The local agent expects a workspace config endpoint that returns a JSON workspace configuration.
With --base-url https://your-control-plane.example.com --workspace mcp-hub, the agent requests:
https://your-control-plane.example.com/v1/workspaces/mcp-hub/config
With export profiles, the control plane can provide separate config URLs for different MCP server views.
Current Scope
mcp-hub focuses on local MCP aggregation and configuration delivery. It is not intended to be a multi-tenant SaaS backend out of the box.
The current implementation is best suited for:
- personal MCP workspaces,
- small team tool setups,
- self-hosted control planes,
- experiments with hosted or generated MCP sources.
Roadmap
- Improve the public setup wizard and install documentation.
- Expand hosted-source lifecycle management.
- Add more automated coverage around source discovery and export profiles.
- Refine the web console for clearer source, tool, and export workflows.
License
MIT
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。