Gateways are All You Need — Karan Sampath, Anthropic

Overview

• Speaker: Karan Sampath
• Channel: AI Engineer
• Main topic: MCP (Model Context Protocol) Gateways in Enterprise Environments
• Purpose: To explain enterprise pain points around MCP server adoption and present gateways as the necessary architectural primitive for secure, scalable AI agent deployments. Karan Sampath from Anthropic discusses the challenges enterprises face when deploying Model Context Protocol (MCP) servers, such as observability, access control, and security bottlenecks. He argues that registries alone are insufficient and introduces the concept of MCP Gateways as the ultimate enterprise control plane to solve these issues, enabling decoupled agent architectures, robust authentication, role-based access control, and secure server deployments.

Topic Map

Introduction and Background

• Explanation: Karan Sampath introduces himself as an Applied AI Forward Deployment Engineer at Anthropic outside the US, working with enterprises on MCP and internal use cases.
• Key claims:
  • Enterprises face severe friction when deploying MCP servers.
• Examples:
  • Working with enterprises on MCP and internal use cases.
• Terminology:
  • MCP
  • Applied AI
  • Forward Deployment Engineer
• Why it matters: Establishes speaker context and industry relevance.

Enterprise State of Play with MCPs

• Explanation: Overview of MCP adoption: open standard created by Anthropic, thousands of servers in the official registry. However, enterprises struggle with three major table stakes: Observability, Access Control, and Security.
• Key claims:
  • Registries help with discovery and installation but are insufficient for enterprise requirements.
  • Observability, access control, and security form a three-headed hydra of enterprise bottlenecks.
• Examples:
  • Who is using my MCP? Which servers and tools are being used?
  • Ensuring correct users have access to each server and correct scoping of tools.
• Terminology:
  • MCP registry
  • Observability
  • Access Control
  • RBAC
  • Security posture
• Why it matters: Highlights why enterprises get stuck and stay with a tiny handful of MCP servers.

The Security Review Bottleneck

• Explanation: Every team builds MCP servers, but they hit security review bottlenecks because security teams are overloaded and lack visibility into tool usage and safety.
• Key claims:
  • Most enterprises stay with a handful of MCP servers due to security review friction.
  • Security teams are unable to see which MCPs teams want to use or verify their safety.
• Examples:
  • Multiple department MCPs feeding into a single security review pipeline.
• Terminology:
  • Security review
  • Bottleneck
  • Server sprawl
• Why it matters: Explains why AI agent deployments stall in large organizations.

What is an MCP Gateway?

• Explanation: Definition of an MCP Gateway as an enterprise control plane sitting between MCP clients and MCP servers, handling auth, authz, ops, deployment, and connectivity.
• Key claims:
  • Gateways serve as a root of trust for security teams.
  • Team-specific business logic is decoupled from security, auth, and network infrastructure.
• Examples:
  • Gateway sitting between MCP clients and servers.
• Terminology:
  • Gateway
  • Control plane
  • Proxy router
  • Auth handler
  • RBAC engine
  • Tunnel handler
  • Subregistry
• Why it matters: Provides a scalable architectural pattern for enterprise AI.

What a Gateway Gives You

• Explanation: Detailed breakdown of gateway benefits: plug-in identity providers once, centralized governance and access control, observability and audit logs, pluggable credentials, and scalability.
• Key claims:
  • Plug into your IDP once, not per server.
  • Centralized audit logging of every tool call and usage metrics across teams and servers.
• Examples:
  • Allowlists, version pins, RBAC, central audit logs.
• Terminology:
  • IDP
  • SSO
  • Delegated user identity
  • Audit log
• Why it matters: Demonstrates the tangible ROI of adopting a gateway layer.

Separating the Agent Harness from Data

• Explanation: The big picture vision: decouple the agent harness (clients) from enterprise data and internal MCP servers via a gateway, enabling agents to run anywhere while data stays secure.
• Key claims:
  • Gateway is the root of trust; agents run anywhere while private MCP servers and data stay inside the network.
  • Separate agents from deployments.
• Examples:
  • Claude managed agents connecting through MCP gateway to enterprise servers.
• Terminology:
  • Agent harness
  • Root of trust
  • VPC
• Why it matters: Enables safe scaling of AI agents without exposing underlying data infrastructure.

Key Points

Registries are necessary but insufficient for enterprises

• Explanation: While registries like modelcontextprotocol.io help with public discovery, they lack authentication, SSO, RBAC, audit logs, and credential management.
• Evidence: Official registry contains thousands of servers, but enterprise adoption stalls on governance.
• Practical implication: Enterprises must build or adopt a control plane beyond basic registries.

The Security Bottleneck restricts AI velocity

• Explanation: Without standardized gateways, every custom MCP server requires individual security reviews, causing massive bottlenecks.
• Evidence: Organizations stay restricted to a handful of approved MCP servers.
• Practical implication: Decentralized tool creation fails without centralized security guardrails.

Gateways establish a Root of Trust

• Explanation: A gateway acts as a single integration point where security teams can bless one platform, allowing individual teams to ship business logic independently.
• Evidence: Gateway handles auth, routing, tunneling, and subregistry management.
• Practical implication:

Frameworks, Models & Processes

MCP Gateway Architecture

• How it works: Sits between MCP clients and MCP servers to intercept, authenticate, authorize, route, and audit all tool calls and server connections.
• Components:
  • Auth Handler
  • RBAC Engine
  • Proxy Router
  • Tunnel Handler
  • Subregistry
  • Tooling CLI
• When to use: When deploying AI agents across enterprise environments requiring strict data governance and multi-team tool development.

Examples & Case Studies

Legal teams building custom MCP servers for contract review

• Illustrates: Enterprise teams wanting to leverage AI tools on internal documents
• Lesson: Legal teams should only care about contract review business logic, not auth, proxying, or security hardening.

Actionable Takeaways

• Immediate:
  • Do not hand-roll ad-hoc MCP security controls per server.
  • Evaluate existing open source or managed gateway providers.
• Strategic:
  • Establish a root of trust using an enterprise gateway layer.
  • Separate the agent harness from your private data infrastructure.
• Questions to investigate:
  • How does our organization handle authentication across multiple AI client applications?
  • What is our current security review process for new tool definitions?

Claims Worth Verifying

• Official MCP registry contains over a thousand servers (statistic)
• Gateways remove the need for individual security reviews per MCP server (architectural claim)

Notable Quotes

"These are important papercuts. The protocol is incredibly powerful — But we must invest in better infrastructure to realize its value." (at 8:41) "Don't hand-roll MCPs. Establish a Root of Trust." (at 8:55)

Compressed Summary

• MCP registries provide discovery but lack enterprise security and governance features.
• Security reviews create severe bottlenecks for enterprise MCP deployments.
• Gateways act as a central control plane for authentication, RBAC, routing, and audit logging.
• Separating agent harnesses from enterprise data via gateways enables secure, scalable AI deployment.
• Keywords: mcp, gateway, security, rbac, governance, agents
• Core insight: Enterprise AI agent deployment requires an MCP Gateway to decouple agent development from security, authentication, and data governance infrastructure.