MCP Servers

Connecting one assistant to an existing service may require only a small adapter, but that work can multiply as more applications need access. A support team might expose ticket lookup to one assistant, then add a second assistant and an editor integration. If each builds its own discovery format, invocation adapter, and result handling, the ticket API stays the same while the surrounding integration work keeps growing.

The Model Context Protocol (MCP) provides a shared interface for exposing capabilities and context to AI applications.

This chapter explains how an MCP server fits around existing services, how clients discover and invoke its features, and which transport, authorization, and operational decisions remain your responsibility.

The wire examples target protocol revision 2026-07-28. That revision uses per-request metadata rather than the initialization handshake that 2025-11-25 and earlier revisions use. Mixing those two models produces incompatible implementations, so version assumptions are explicit throughout.