Project · AI tooling
MCP Hub
A shared gateway for tool discovery, authentication, routing and inspection across MCP servers.
- AgentAuthenticated MCP client
- HubResolve registrations and tool names
- Request checksContent safety, DLP and session checks
- Policy blockReturn an error
- Upstream callExecute, then inspect response
One connection to a tool catalogue
I built a hub that presents tools from multiple Model Context Protocol servers through one endpoint. It resolves the caller’s registrations and credentials, routes each call to the appropriate upstream server, and applies the configured inspection policy.
Namespaced tool names keep upstream tools distinguishable. A tool cache and a Redis-backed circuit breaker help manage repeated discovery and failing upstream services.
Policy in both directions
The recorded request pipeline runs content-safety inspection, DLP filtering and a session-buffer check before the upstream call. The response passes through DLP and content-safety inspection before being returned. Call metadata records latency, status, warnings and filter outcomes.
Policies are configurable per user and require explicit enablement. The implementation notes describe fail-open scanner errors: exceptions are logged and the call continues. Explicit detections in block mode stop the relevant request or response. This behaviour is a material part of the gateway’s enforcement model.
Related work
The hub connects the DLP pipeline with hosted MCP sandboxes. Each layer has its own responsibility for inspecting data, routing calls or containing execution.