Uber Builds MCP Gateway Spanning 800 Servers and 5000 Tools

By 作者:Leela KumiliOriginal title: 800 个 MCP Server、5000 个工具,AI 智能体的“微服务危机”来了
78AI Score

Uber has built a unified MCP Gateway that governs over 800 MCP servers and 5,000 tools, separating a Registry control plane from a Proxy runtime so existing APIs can be exposed as MCP tools without modifying underlying services. The design is widely read as the API gateway pattern reborn for the agent era, while surfacing familiar distributed-systems problems such as discovery, authorization and observability at tool scale.

Full Report

As AI agents became widely adopted within the company, Uber built an MCP Gateway to standardize how agents discover and invoke internal services, addressing issues such as fragmented integrations, duplicated infrastructure, and inconsistent security practices. The platform supports over 800 MCP servers and more than 5,000 tools, providing a unified layer for service discovery, protocol translation, authorization, and execution.

Uber engineers describe the solution as a microservices architecture that can expose existing APIs as MCP tools without modifying the underlying services. The overall architecture separates the MCP Registry, which serves as the control plane, from the Proxy Gateway, which handles runtime execution. The Registry maintains tool definitions, schemas, ownership, and enablement settings; the Proxy translates MCP requests into HTTP, gRPC, or TChannel calls, then converts responses back into MCP-compatible results.

The architecture has been compared to the well-established API gateway pattern. Ajay Sharma, a Principal Software Engineer at Microsoft CoreAI, responded in a LinkedIn post:

Azure API Gateway has been doing this for a long time. This is a way to reorganize the chaos brought by large language models into order.

Kelly Goetsch, President of Pipe17, also noted in a LinkedIn post:

Uber's new MCP gateway looks exactly the same as the gateways we've been building for APIs, and later for GraphQL, over the past few decades.

Tamir Dresher emphasized in a LinkedIn post that traditional distributed systems concerns such as security, service discovery, observability, and governance remain just as important today.

Uber's AutoCrawler uses Cadence to automate tool discovery. It scans internal interface definition language (IDL) registries, extracts API definitions from Protocol Buffers and Thrift schemas, then generates MCP-compatible schemas and uses large language models to generate tool descriptions. Newly discovered tools remain disabled by default until their owning teams complete review and explicitly enable them. Configuration changes can propagate directly to the data plane without restarting services or redeploying.

The gateway enforces authorization at both the server and tool levels through Uber's internal access control system. It also supports redaction of personally identifiable information (PII) and other sensitive data in tool responses. For third-party integrations such as Jira and Google, a separate service exchanges internal user tokens for the corresponding third-party authentication tokens.

As the tool catalog continues to expand, loading definitions for thousands of tools into the agent context becomes inefficient. Uber developed Omni MCP for this purpose, supporting progressive discovery that lets agents search servers, discover relevant tools, retrieve individual schemas, and invoke selected tools on demand. Response Projection allows agents to specify the response fields they need rather than receiving the full payload. Uber's aifx command-line interface also supports coding agents in discovering and invoking tools, and writes results to files for selective viewing.

Original link:

https://www.infoq.com/news/2026/11/uber-mcp-gateway-ai-agents/