Uber introduced an MCP Gateway that standardizes discovery and invocation of internal services for AI agents across a fleet of over 800 MCP servers and 5,000 tools. By separating tool registration from runtime execution and adding automated API discovery, access controls, sensitive‑data redaction and response filtering, the platform gives engineers a unified entry point for agent‑service interactions.
Architecture Shift
The gateway sits between agents and the underlying services, acting as a broker that resolves service definitions and forwards calls. Registration of a tool no longer requires agents to embed endpoint details; instead, tools are cataloged once in the gateway and become discoverable to any authorized agent. Incremental discovery is supported through Omni MCP, allowing new tools to appear without redeploying agents.
Operational Impact
Deploying a central gateway introduces a new scaling consideration: the gateway must handle the aggregate traffic of thousands of agents and services. Operators will need to monitor registration latency, API‑discovery refresh cycles, and the health of the gateway itself. Because registration is decoupled from execution, CI/CD pipelines can push tool definitions independently of agent releases, but they must also ensure that registration data stays consistent with runtime expectations.
Security Considerations
Built‑in access controls let the gateway enforce which agents may invoke which tools, while sensitive‑data redaction strips confidential fields from responses before they reach the agent. Response filtering provides an additional layer to limit the surface of returned data. Practitioners must configure these controls deliberately, as they become the primary gatekeeper for data exposure and misuse.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Teams should evaluate how their existing agent code integrates with a brokered discovery model, update registration workflows to target the gateway, and establish monitoring for gateway latency and error rates. Security owners need to review the default access‑control policies and redaction rules to align with data‑handling requirements. Finally, plan incremental rollout to validate that the gateway’s discovery and filtering behave as expected before deprecating any legacy direct‑call mechanisms.
