MCP — Model Context Protocol intro

If every AI application (Claude Desktop, an IDE plugin, a corporate assistant) needs to integrate with every enterprise system (Jira, Salesforce, Postgres, GitHub) individually, you get an N × M explosion of custom integrations. MCP is a standard protocol that turns this into N + M: each application speaks MCP, each system exposes an MCP server, and any app can use any system.

Before USB-C: every device had its own connector, and you needed the right cable for each. USB-C standardized one connector so any charger works with any device. MCP is USB-C for AI: one standard interface between AI applications and external systems. Ship an MCP server for your system once, and every MCP-compatible AI app can use it.

An MCP server exposes three things. Tools: functions the AI can call (like the function calling from D0L4). Resources: data the AI can read (files, query results, URL contents). Prompts: reusable prompt templates users can invoke. Most servers focus on tools, but resources and prompts are equally valid primitives.

Host = the AI application the user interacts with (Claude Desktop, Claude Code). Client = a connector inside the host that speaks MCP. Server = the external process that exposes tools/resources/prompts. One host can host many clients, each connected to a different server. Communication is via JSON-RPC over stdio (local) or streamable HTTP (remote).

Claude Desktop (host)
 ├── client A ↔ Filesystem MCP server (stdio, local)
 ├── client B ↔ Postgres MCP server (stdio, local)
 └── client C ↔ Jira MCP server (HTTP, remote)

Because it turns 'integrate our AI assistant with all our systems' from a bespoke engineering project into a standardized deployment problem. Vendors ship MCP servers for their products. Enterprises deploy them, configure auth, and every MCP-compatible AI in the org can talk to those systems consistently, with audit logging and access control at the server layer.

You don't need to memorize the JSON-RPC spec. You need to know: the architecture (host/client/server), the three primitives, when to use local vs remote deployment, the security model, and how tool design principles apply to MCP tools too. D4 covers all of this in depth.