AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Uber Builds MCP Gateway to Govern Access to 800+ Servers and 5,000+ AI Agent Tools

Collected Nov 2, 2026

Uber has built an MCP Gateway to standardize how AI agents discover and invoke internal services, according to a report by Leela Kumili. The platform now supports more than 800 MCP servers and 5,000 tools, and provides a shared layer for discovery, protocol translation, authorization and execution. Uber's stated motivation is to counter fragmented integrations, duplicated infrastructure and inconsistent security practices as agent adoption spreads across the company.

Model Context Protocol, or MCP, is an open standard that lets an AI agent call external tools and data sources through a common interface. In practice, each MCP server publishes a set of tools with schemas describing their inputs and outputs; the model reads those descriptions and decides which tool to invoke. Once many teams each run their own servers, the catalog becomes hard to govern, which is the problem Uber says its gateway addresses.

Uber engineers describe the approach as a microservice architecture that exposes existing APIs as MCP tools without requiring changes to the underlying services. That is a meaningful constraint: rather than asking service owners to rewrite endpoints for agents, the gateway adapts what already exists.

The architecture splits into two halves. The MCP Registry acts as the control plane, holding tool definitions, schemas, ownership records and enablement settings. The Proxy Gateway handles runtime execution: it translates incoming MCP requests into HTTP, gRPC or TChannel calls, then converts the responses back into MCP-compatible results. TChannel is Uber's own RPC protocol, so the gateway reaches services speaking the protocols Uber already runs internally rather than only modern ones.

Tool discovery is automated by a component called AutoCrawler, built on Cadence, Uber's workflow orchestration system. AutoCrawler scans internal interface definition language registries, pulls API definitions out of Protocol Buffers and Thrift schemas, generates MCP-compatible schemas, and uses a large language model to write the tool descriptions that agents read. Newly found tools stay disabled until their owning teams review them and explicitly switch them on. Configuration changes reach the data plane without service restarts or redeployments, which suggests the registry pushes updates to the proxy rather than requiring a rollout per change.

Authorization is enforced at two levels, server and tool, through Uber's internal access control system. The gateway also redacts personally identifiable information and other sensitive data from tool responses. For third-party integrations such as Jira and Google, a separate service swaps internal user tokens for the matching third-party authentication tokens, so agents inherit the calling user's external permissions rather than a shared credential.

Scale creates a context problem. Loading definitions for thousands of tools into an agent's context window is inefficient and expensive, and it can degrade model accuracy by crowding out the actual task. Uber's answer is Omni MCP, which supports incremental discovery: an agent searches for servers, finds relevant tools, retrieves individual schemas and invokes only what it selects. A companion feature, Response Projection, lets agents request specific fields from a response instead of receiving the entire payload. Uber's aifx command-line interface lets coding agents discover and invoke tools and write results to files for selective inspection.

The report notes that practitioners compared the design to established API gateway patterns. Ajay Sharma, a principal software engineer at Microsoft CoreAI, wrote on LinkedIn that Azure API Gateways has been doing this for a while and called it a recipe for order amid the chaos LLMs bring. Kelly Goetsch, president at Pipe17, said the Uber gateway looks identical to gateways built over the decades for APIs and later GraphQL. Tamir Dresher pointed on LinkedIn to the continuing relevance of distributed systems concerns such as security, discovery, observability and governance. My reading is that this comparison is fair but incomplete: the control-plane and data-plane split, token exchange and schema generation do mirror API gateway practice, yet tool descriptions written by a model and an LLM deciding which tool to call add failure modes traditional gateways never faced, such as misleading descriptions or over-broad tool selection.

Several trade-offs follow. Requiring owner approval before a tool goes live slows catalog growth but puts accountability on the team that owns the API. Redaction at the gateway adds a safety net but is only as good as the patterns it applies. Incremental discovery reduces context load but adds round trips, since an agent must search and fetch schemas before acting. And exposing existing APIs unchanged keeps maintenance low while passing any quirks of those APIs through to agents.

Why it matters: Teams building agent platforms face the same trajectory Uber describes, where a handful of MCP servers becomes hundreds and security and discovery turn into bottlenecks. The split between a registry and a runtime proxy, plus approval gated tool registration and per-tool authorization, is a reusable blueprint for that stage; my inference is that other large enterprises will converge on similar gateways, possibly by extending existing API management products rather than building new ones. Product and platform teams planning agent integrations should assume governance, not just connectivity, becomes the core work.

Read at InfoQ · AI, ML & Data Engineering

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Uber has built an MCP Gateway to standardize how AI agents discover and invoke internal services across more than 800 MCP servers and 5,000 tools. The gateway separates tool registration from runtime execution and includes automated API discovery, access controls, sensitive data redaction, incremental tool discovery through Omni MCP, and response filtering. By Leela Kumili