Module 1: MCP (Model Context Protocol)
The first problem any AI agent hits: it needs data from multiple systems, and every system speaks a different language.
The Problem
Your contract processing agent needs to pull documents from email, check vendor records in the CRM, retrieve company policies from SharePoint, and look up past analyses in a document database. Each source has its own authentication, data format, and API quirks.
Without a standard, you're writing custom integration code for every data source. That's fragile, hard to test, and painful to maintain.
What MCP Solves
MCP (Model Context Protocol) is an open standard that gives AI agents a single, unified way to connect to any external system. Instead of teaching the agent how to talk to each data source individually, you define an MCP server for each source and the agent connects to all of them the same way.
Think of MCP as USB for AI agents. One standard. Everything plugs in.
How It Works
An MCP server exposes three types of capabilities:
| Capability | What It Does | Example |
|---|---|---|
| Resources | Read-only data the agent can access | Company policy documents, vendor records |
| Tools | Functions the agent can call | Search email, query database, generate PDF |
| Prompts | Pre-built prompt templates | Contract analysis template, risk assessment template |
The agent discovers what's available through the MCP protocol, then uses what it needs for the current task.
In Our Contract Workflow
Email Inbox ──→ MCP Server ──┐
CRM ──→ MCP Server ──┤──→ Contract Agent
SharePoint ──→ MCP Server ──┤
Doc Database ──→ MCP Server ──┘
Each MCP server is a thin wrapper around the underlying system. The agent doesn't know or care whether it's talking to an email API, a REST endpoint, or a database. It just knows it has resources to read and tools to call.
Key Concepts
- Server discovery - The agent finds available MCP servers at startup
- Capability negotiation - Each server advertises what it can do
- Stateless protocol - Each request is independent, making it easy to scale
- Security boundaries - Each server controls its own authentication and access
Why This Matters in Production
Without MCP, adding a new data source means updating the agent's code, testing the integration, and redeploying. With MCP, you deploy a new MCP server and the agent discovers it automatically. The agent's code doesn't change.
This is the difference between a demo that talks to one database and a production system that connects to your entire enterprise.
What's Next
MCP gives the agent access to data. In Module 2: Tool Use, we cover how the agent takes action on that data.
MCP Implementation Lab
Build MCP servers for S3, DynamoDB, and a REST API. Wire them to a Bedrock agent with full authentication and error handling.