AI agents are evolving into primary interfaces for enterprise systems, necessitating robust security measures beyond simple API keys. This article delves into implementing an identity-aware API gateway for Model Context Protocol (MCP) tools, utilizing a triple-token model: user (OIDC), agent (OIDC-A), and tool/resource tokens. It also explores fine-grained RBAC policies, session-affinity routing, and centralized auditing. Feedback from the identity, security, and AI-agent communities is welcome.