Uber's MCP Gateway is the most useful enterprise AI story of the month, and it has almost nothing to do with model capabilities. The company built a control plane that standardizes how AI agents discover and invoke internal services across more than 800 MCP servers and 5,000 tools. That is the unglamorous plumbing that decides whether agents work in production or stall in a demo. Anyone who has watched an agent architecture succeed on a laptop and fail behind a corporate firewall should recognize what Uber is solving. The same discipline shows up in our look at Six Architectures for Smarter Enterprise AI Beyond Basic RAG, where the hard work is engineering precise, cost-efficient context rather than chasing a bigger model. Uber is making the same bet at the infrastructure layer.
The design choices are the real signal. Separating tool registration from runtime execution means an agent does not need a live connection to every service just to know what exists. Automated API discovery and incremental tool discovery through Omni MCP keep the catalog current without turning it into a maintenance burden. Access controls, sensitive data redaction, and response filtering sit at the gateway rather than inside each agent, which is how governance stays consistent as the number of servers and tools grows. This is what mature AI operations look like: the agent gets simpler, and the platform absorbs the complexity. It mirrors the practical thinking behind AI-Native Spreadsheets: Smarter Data Management Without the Learning Curve, where the goal is to remove friction for the people doing the work rather than to impress them with machinery.
Our take is direct: governance at the gateway is the pattern enterprises should copy, and the tool catalog is the asset that matters. Five thousand tools is not a benchmark to brag about. It is a surface area that needs permissions, redaction, and filtering to stay safe, and Uber has put those controls in one place instead of scattering them across teams. The separation of registration from execution also gives platform teams room to change backends without breaking every agent that depends on them. That is the kind of decision that saves quarters of rework later.
The open question is whether this pattern travels. Uber can staff a platform team to run a gateway across 800 servers; a mid-sized company may not be able to. Watch for whether gateway capabilities get packaged into something smaller organizations can adopt without building their own. The specific consequence to track is simpler: once discovery, access control, and redaction are standardized, adding a new tool to an agent becomes a registration task rather than an integration project. That is the moment agents start scaling inside a business instead of stalling at the pilot stage.