Skip to content
Adaptive Agility

3.4 — MCP (Model Context Protocol)

What Is MCP?

Section titled “What Is MCP?”

MCP (Model Context Protocol) — a standard protocol (set of rules) that lets AI agents connect to external tools and data sources in a consistent way.

Analogy: USB is a standard that lets any device plug into any computer. MCP is a standard that lets any AI connect to any tool. Before MCP, every tool connection was custom-built — one-off integrations that only worked with one harness. MCP standardizes it so the work gets done once and everyone benefits.

How It Works

Section titled “How It Works”
AI Agent ←→ MCP Protocol ←→ MCP Server ←→ External System
                                           (database, API,
                                            Slack, GitHub,
                                            file system, etc.)

The AI never talks directly to the external system. It talks to the MCP protocol layer, which routes the request to the right MCP Server, which handles the translation.

Three Components

Section titled “Three Components”

MCP Server — a small program that translates between the AI and an external system. It exposes “tools” the AI can call. Someone builds an MCP Server for Slack once; any AI harness can then use it.

MCP Client — the AI’s side of the connection. Built into harnesses like Claude Code and Cursor. You don’t build this — your harness provides it.

MCP Tool — a specific action the server makes available. Examples:

  • Read a Slack message
  • Query a database
  • Create a GitHub issue
  • Look up a calendar event
  • Search a knowledge base

Each MCP Server exposes multiple tools. Each tool has a name and a defined set of inputs — the AI knows what it can call and what parameters each tool takes.

Why This Matters for Builders

Section titled “Why This Matters for Builders”

MCP lets you give your AI access to your tools, your data, your systems — without writing complex integrations from scratch.

Before MCP, connecting an AI to your database required building a custom integration specific to your harness. With MCP, you install or write an MCP Server, and any MCP-compatible harness (Claude Code, Cursor, and others) can use it.

Practical implications:

  • Thousands of MCP Servers already exist for common tools (GitHub, Slack, databases, web search, and more)
  • You can install them in minutes and give your agent new capabilities immediately
  • If no server exists for what you need, you can build one — or hire someone to build one — and reuse it everywhere

This is the tool layer from the Agent Stack made concrete. Adding an MCP Server is how you expand what your agent can do.

In Phase 4, you’ll see how MCP fits into larger multi-agent architectures where specialized agents each have access to different tool sets.

Next: 3.5 — Memory & RAG | Phase overview: Phase 3