MCP as the New Internal API Layer

▶ Watch (00:52)

MCP (Model Context Protocol) was built by Anthropic to give Claude context from external systems. It started as local Mac access and expanded to remote servers. Where a CLI takes typed commands and runs each in isolation, MCP connects tools to an LLM so the model can reason, chain calls, and make decisions across a full workflow. A tool description like “manages feature flags for customers” is what the LLM reads to decide whether to call it. That natural-language contract is the interface.

Three Enterprise Blockers: Auth, Registries, Audit Logs

▶ Watch (04:32)

MCP was announced in November 2024 and is not yet a year old. That youth shows in what’s still missing for enterprise use: auth, server discovery, and audit trails. The auth requirement is OAuth 2.1, not OAuth 2.0. The difference is real: 2.1 adds dynamic client registration, where the server generates a client ID and secret in real time. A registry launched in September 2025 but is a bare index, not a curated store. Audit logs are non-negotiable. In FINRA-regulated environments, Grinich noted, missing logs means legal exposure.

Build Guidelines That Keep MCP Production-Safe

▶ Watch (10:34)

Grinich’s core lesson: design tools around what people do, not around your API surface. Auto-generated MCP servers from OpenAPI specs bloat the context window and degrade model attention. Fewer tools means better performance. Each MCP server should be scoped to one specific action, documented for both human developers and the agents that will call it programmatically, and gated behind an approval flow. He walked through the Live MCP shirt-ordering demo via ChatGPT (12:47) to show dynamic client registration connecting a merch store server to ChatGPT over OAuth.

Notable Quotes

Clear use cases. You do not want to have one tool per API endpoint. Michael Grinich · ▶ 10:53

Second here, don’t don’t ignore best practices around security. Michael Grinich · ▶ 11:47

you don’t want to just open the floodgates to everyone, Michael Grinich · ▶ 12:22

Key Takeaways

  • Map MCP tools to user actions, not API endpoints, to keep LLM context lean and performance high.
  • MCP auth requires OAuth 2.1 with dynamic client registration — OAuth 2.0 is not a drop-in substitute.
  • Scope each agent to one specific action, log every interaction, and gate who can connect to which server.