# Uber builds MCP gateway to govern agent access to 800+ servers

published: 2026-11-02T14:05:28.811Z
topic: mcp
source: https://www.infoq.com/news/2026/11/uber-mcp-gateway-ai-agents/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
url: https://aiagentnews.io/stories/uber-builds-mcp-gateway-to-govern-agent-access-to-800-servers

Uber has built an internal MCP Gateway that lets AI agents discover and call its services through one shared layer, covering more than 800 MCP servers and 5,000 tools. The design adds central authorization, PII redaction and incremental tool discovery on top of existing APIs without changing the services behind them.

Uber built the gateway to standardize how agents discover and invoke internal services, as fragmented integrations and inconsistent security grew with agent adoption. A registry holds tool definitions, schemas, ownership and enablement settings, while a proxy translates MCP requests into HTTP, gRPC or TChannel calls. Existing APIs are exposed as MCP tools without changes to the underlying services.

AutoCrawler scans IDL registries for Protocol Buffers and Thrift definitions and generates MCP schemas, using a large language model to write tool descriptions. New tools stay disabled until their owning teams enable them. Authorization applies at server and tool levels, responses can have PII redacted, and a separate service exchanges tokens for third-party services such as Jira and Google.

Omni MCP supports incremental discovery, so agents search for servers, fetch individual schemas and invoke only the tools they need rather than loading thousands of definitions. Response Projection returns only requested fields, and the aifx command-line interface lets coding agents write results to files for selective inspection.

## Why it matters if you build agents

If you build agents that call internal APIs, the pattern shows how to expose existing services as MCP tools through a registry, per-tool authorization and opt-in enablement instead of wiring each integration separately. Loading thousands of tool schemas into context is costly, so incremental discovery and field-level response projection are worth copying. Several practitioners noted the design resembles long-standing API gateways, so the main value is governance rather than novelty.