Uber MCP Gateway Turns Thousands of Internal APIs into Agent Tools
The non-obvious takeaway is that scaling agents inside a large enterprise is a governance and connectivity problem, not a model problem — Uber solved it by retrofitting thousands of existing services rather than rewriting them. Engineering teams building internal agent platforms can copy the progressive-discovery pattern and the discovery-is-not-exposure security model. Compared with ad hoc per-team MCP integrations, this centralised gateway approach offers a reusable blueprint for any organisation with a large legacy API surface.
Uber has deployed an internal MCP gateway built on a control plane plus data plane architecture, currently hosting 800+ MCP servers and 5,000+ tools.
Its AutoCrawler pipeline, running on the Cadence workflow engine, scans internal Protobuf/Thrift IDL registries and uses an LLM to generate agent-friendly tool descriptions, converting legacy APIs to MCP JSON-RPC 2.
0 without any code changes from service teams.A registry serves as the single source of truth, with all tools disabled by default and enabled only after owner review, plus per-tool authorization and automatic PII masking.
To fight context bloat, an Omni MCP entry point with four meta-tools enables progressive tool discovery, while response projection trims payloads at runtime.
Uber MCP 网关设计工程实践:面向 AI Agents 的企业级 MCP 管理平台 @UberEng Uber 把全公司数千个存量 API 用一套 “控制平面 + 数据平面” 的网关体系自动转译成 MCP 工具,让 AI Agents 不改动任何后端服务就能调用整个公司的能力。目前已托管 800+ MCP 服务器、5000+ 工具。 要解决的问题 AI Agents 在 Uber 内部快速铺开,数百个团队各自搭建与后端服务的 MCP 集成,导致工具碎片化和基础设施重复建设。Uber 的判断是:缺的不是更聪明的模型,是一套统一的“连接组织”,发现、安全、可靠性这些让 Agents 值得信任的基建。 架构:经典的两平面分离 · MCP Registry(控制平面):全公司工具的目录和唯一事实来源,管理发现、所有权、启停。支持从“零代码包装现有 API”到“原生 MCP 服务”的多种形态。 · Proxy Gateway(数据平面):运行时真正执行 MCP 请求,在 MCP 协议与 HTTP / gRPC / TChannel(Uber 自研 RPC 协议)之间双向转换,经 Muttley(服务网格 sidecar)转发到下游,再把响应转回 MCP JSON。配置热更新,无需重启。 最有工程价值的三个设计 1. AutoCrawler:存量 API 的自动化“MCP 化” 基于 Cadence 工作流引擎,持续扫描内部 IDL 注册表(Protobuf/Thrift 定义),对每个新服务执行五步流水线:建虚拟服务器 → 解析 schema 和文档注释 → 用 LLM 生成对智能体友好的工具描述 → 把 Protobuf/Thrift 转成 MCP 兼容的 JSON-RPC 2.0 schema → 注册。整个流程不需要服务团队写一行代码。 2. “发现 ≠ 暴露” 的安全模型 所有工具默认禁用,必须由拥有该服务的团队显式审核后启用;工具描述的变更要走配置 diff 审批、可回滚。授权细到单个工具,按调用主体(人/服务/Agents)应用不同策略,PII 数据开箱即用自动脱敏。这是把平台工程的传统治理纪律完整搬进了智能体时代。 3. 对抗上下文膨胀的三板斧 给智能体配几百个服务器会直接撑爆模型上下文窗口,Uber 的解法层层递进: · Omni MCP:单一入口 + 四个元工具(discover_server → discover_tools → get_tool_schema → invoke_tool),让智能体按需渐进式发现,而不是一次性加载全部定义; · Response Projection:类 GraphQL 思路,让模型在请求里声明只需要哪些字段,网关在运行时裁剪响应; · Code Mode:编码智能体通过 aifx CLI 调用工具并把输出写入文件,再用 grep 选择性读取;彻底绕开上下文限制,目前已是 Uber 内部编码智能体的默认方式。 Uber Engineering @UberEng x.com/i/article/2106… 🔗 View Quoted Tweet📊 3 ⚡
This report is categorized under AITOP's "Products" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include MCP and tools, agents, Uber, API gateway.
Source attribution: curated from shao__meng (@shao__meng).