# How Kong Runs a Secure, Cost-Efficient AI Gateway with Just Two Engineers

This talk explains how Kong’s two-person AI platform team manages AI gateway access for both production services and internal users at scale. Suprit describes a dual-gateway approach: one gateway supports production applications such as chatbots, search, alert triage, and the Kai platform, while another serves engineers, sales, and internal agents that need faster access to a broader set of models. Instead of distributing API keys, the team uses OIDC and enterprise identity providers like Okta for authentication, central ACLs for authorization, and role assumption to reach Bedrock without storing model credentials in the gateway or client side. The setup also proxies MCP tools such as Slack, GitHub, Atlassian, and Datadog through a token vault, keeping upstream secrets protected. Key operational practices include cost attribution by product and user, regional routing, guardrails, rate limiting, bot detection, and detailed observability for usage, failures, uptime, latency, and prefix cache performance. The team supports roughly 45 models, with 1.4 million requests and 53 billion tokens processed in a recent week.