Descriptions:
Karan Sampath, a forward-deployed engineer at Anthropic and the company’s first such hire outside the US, presents Anthropic’s framework for solving enterprise MCP (Model Context Protocol) deployment challenges at the AI Engineer conference. The talk, framed around the thesis ‘gateways are all you need,’ argues that while the official MCP registry now hosts thousands of servers and teams are rapidly building more, enterprises consistently hit the same wall: observability, access control, and security are completely opaque in current MCP deployments.
Sampath’s proposed solution is an MCP gateway layer—a middle tier handling authentication, role-based access control, routing through a secure proxy, encrypted tunnel connections, an internal server sub-registry, and a gateway CLI. Under this architecture, individual MCP servers only need to implement business logic; the gateway absorbs all infrastructure concerns. A legal team building a contract review server, for example, can focus entirely on redlining logic rather than access management, scalability, or auditing.
The talk draws a direct parallel to how API infrastructure solved analogous problems in the prior generation of enterprise software—enterprises already know how to secure an API endpoint, but equivalent patterns for MCP are still being established. Sampath frames the gateway as foundational infrastructure for multi-agent deployments, enabling AI clients like Claude Code to discover and invoke internally-hosted tools through a standardized, auditable layer. The session includes a detailed walkthrough of gateway components and the specific security threat model that current MCP deployments lack.
📺 Source: AI Engineer · Published April 27, 2026
🏷️ Format: Deep Dive
