Uber has deployed an MCP Gateway that serves as the governed front door for AI agents accessing internal services. The platform currently supports more than 800 MCP servers and over 5,000 tools. It uses a dual-plane design: an MCP Registry acts as the control plane holding tool catalogs, ownership, and enablement state, while a Proxy Gateway handles runtime translation of MCP calls into HTTP, gRPC, or TChannel requests through Uber’s service mesh. An AutoCrawler workflow continuously scans interface definition registries, generates MCP schemas from Protobuf and Thrift definitions, and uses a language model to create tool descriptions. New tools remain disabled until the owning team reviews and enables them. Authorization is enforced at both server and tool levels, with default redaction of personal data. The approach allows existing APIs to be exposed as MCP tools without changes to the underlying services.
Uber has deployed an MCP Gateway that serves as the governed front door for AI agents accessing internal services. The platform currently supports more than 800 MCP servers and over 5,000 tools. It uses a dual-plane design: an MCP Registry acts as the control plane holding tool catalogs, ownership, and enablement state, while a Proxy Gateway handles runtime translation of MCP calls into HTTP, gRPC, or TChannel requests through Uber’s service mesh. An AutoCrawler workflow continuously scans interface definition registries, generates MCP schemas from Protobuf and Thrift definitions, and uses a language model to create tool descriptions. New tools remain disabled until the owning team reviews and enables them. Authorization is enforced at both server and tool levels, with default redaction of personal data. The approach allows existing APIs to be exposed as MCP tools without changes to the underlying services.