MCP server sprawl, meet your new traffic cop

What it does

MCPHub sits between AI clients and a fleet of Model Context Protocol servers, exposing them through a single set of Streamable HTTP (SSE) endpoints. Clients connect to one port and reach either every registered server, a named group, or a specific instance, while an included React dashboard and CLI let you monitor, reconfigure, or call tools without touching each server individually. Configuration starts as a plain JSON file, but you can flip to PostgreSQL for persistence and vector search.

The interesting bit

The “Smart Routing” endpoint uses pgvector to run semantic search over tool descriptions, effectively acting as a tiny AI concierge that guesses which server actually has the tool a client is asking for. That turns a dumb reverse proxy into something that understands intent, which is unusual for infrastructure glue.

Key highlights

• Hot-swappable server registry: add, remove, or update MCP servers without restarting the hub.
• Granular visibility controls inside groups: independently hide or expose Tools, Prompts, and Resources per server.
• Built-in auth stack: local JWT accounts, bearer keys, a full OAuth 2.0 server, and optional GitHub/Google login via Better Auth.
• Headless mode: the backend and MCP endpoints can run without the bundled React UI.
• The same binary serves as a CLI for remote management and marketplace discovery.

Caveats

• Smart Routing and social login are gated behind Database Mode; the default file-based configuration does not support them.
• The project explicitly warns that Windows developers must start the backend and frontend separately, so the monolithic dev server appears to be Unix-biased.

Verdict

Teams running more than one MCP server behind a firewall should look here. If you only need a single local stdio server, this is overkill.