How Uber Built A Gateway For AI Agents To Access Tools
Loading article
Software Architecture/Artificial Intelligence
How Uber Built A Gateway For AI Agents To Access Tools
ByteByteGo·3w ago·14 min read
By ByteByteGo
AI AgentsKubernetesTerraformPrompt EngineeringArtificial IntelligenceSoftware ArchitectureAPI DesignAWS
AI summary
Uber developed an MCP (Model Context Protocol) Gateway to unify how AI agents access internal enterprise tools. The platform provides a centralized layer for discovery, security, and permission management, allowing agents to interact with existing internal services without requiring teams to rewrite their APIs. The architecture includes features like response projection to limit data payloads and code mode to facilitate agentic interactions with files.
Key points
The primary challenge in building agentic systems is creating reliable connective tissue for discovery and security, not the AI models themselves.
The MCP Gateway acts as a centralized middleware layer to solve fragmentation caused by teams independently managing AI integrations.
Leveraging existing APIs through an adapter gateway allows for faster integration without requiring service-level refactoring.
Response projection patterns (similar to GraphQL) help manage token usage and response size by filtering tool result fields.