Uber MCP Gateway – Somasundara & Tehrani
Speakers: Meghana Somasundara (Agentic AI Lead) and Rush Tehrani (Head of Engineering, Agentic AI Platform), Uber
Uber Scale Metrics
| Metric | Value |
|---|---|
| Engineers | 5,000+ |
| Engineers using AI weekly | 90%+ |
| Internal services | 10,000+ |
| Internal agents | 1,500+ |
| Executions per week | 60,000+ |
"This is no longer a pilot program at Uber. It's the new standard for how we work."
MCP Gateway & Registry
Built a control plane for all MCP interactions:
Gateway Orchestrator: Crawls 10,000+ service IDLs, generates MCP tool definitions with LLM, stores in object storage.
Gateway Service: Serves MCP servers, allows service owners to update definitions, diff, security scan, and commit.
Security: Authorization service integration, PII redaction (automatic), periodic code scanning, full observability and audit trails. Third-party MCPs get extra data scanning and sandboxed environments.
Three Consumption Surfaces
| Surface | Type | Key Features |
|---|---|---|
| Uber Agent Builder | No-code | Thousands active monthly, workflow automation |
| Uber Agent SDK | Code-first | Grocery assistant, care coordination, customer support |
| Coding Agents | Developer tools | Claude Code, Cursor, Minions — 1,800 code changes/week, 95% of engineers |
Agent Configuration Patterns
- Agent Builder: Mention MCP server in instructions, pick tools, parameter overlays
- Agent SDK: YAML config with MCP name, tool selection, parameter overrides
- Coding Agents: AIFX CLI —
mcp addcommand for remote/local servers
Roadmap
- Tool Search — Discover tools automatically, load on demand
- Skills — Recipes for using MCPs, shareable across teams
- Evaluations — Evaluate skill invocation, output quality, A/B testing
Key Quotes
"With thousands of MCPs, chaos was coming quickly. MCP and registries were going to make AI usable at numbers."
"Coding agents are producing about 1,800 code changes a week right now, used by 95% of our engineers."