Skip to content

MCP_GUIDE.md

Latest commit

 

History

History
154 lines (103 loc) · 5.76 KB

MCP_GUIDE.md

File metadata and controls

154 lines (103 loc) · 5.76 KB

Understanding the Model Context Protocol (MCP) 🔌

MCP Guide — Part of AgenticAI Foundry
For MIT Professional Education: Applied Generative AI for Digital Transformation

What is MCP?

The Model Context Protocol (MCP) is an open standard that lets AI models talk to external tools and services using a single, universal protocol.

The USB-C Analogy

Think about what happened with device chargers:

Before USB-C With USB-C
Mini USB for cameras One connector for everything
Micro USB for phones Charges your laptop
Lightning for iPhones Transfers files to your phone
Barrel plugs for laptops Connects your monitor
DisplayPort for monitors Works with any brand

Before MCP, every AI model needed a custom integration for every tool it wanted to use. 10 AI models × 10 tools = 100 custom connectors — fragile, expensive, hard to maintain.

With MCP, each AI model implements one standard client, and each tool implements one standard server. 10 AI models + 10 tools = 20 connections. Build once, connect to anything.

How MCP Works

MCP has three layers:

1. AI Model (MCP Client)

The AI model — Claude, GPT, Gemini, etc. — that receives user requests and decides which tools to use.

  • Parses natural language into structured intent
  • Decides which MCP tools to call and in what order
  • Interprets results and responds to the user

2. MCP Protocol Layer

The standardized communication format. All messages use JSON-RPC 2.0 with defined methods:

Method What It Does
tools/list Discover what tools are available
tools/call Invoke a specific tool with parameters
resources/read Access data sources
prompts/get Retrieve prompt templates

3. MCP Servers (Tools)

External services wrapped in a standard MCP interface. Examples:

  • Google Calendar, Outlook
  • Slack, Discord, Microsoft Teams
  • Salesforce, HubSpot
  • GitHub, Jira
  • Datadog, PagerDuty
  • Spotify, file systems, databases

Each server handles its own authentication, rate limiting, and error handling internally. The AI model doesn't need to know those details.

A Real Example: Scheduling a Meeting

User says: "Schedule a meeting with Sarah for next Tuesday at 2pm"

Here's what happens behind the scenes:

Step Component What Happens
1 AI Model Parses intent: schedule_meeting(contact='Sarah', date='next Tue', time='2pm')
2 MCP Client Discovers available tools → finds google_calendar MCP server
3 MCP Protocol Sends JSON-RPC: { method: 'tools/call', params: { name: 'create_event', ... } }
4 Calendar Server Checks Sarah's availability, creates event, sends calendar invite
5 AI Model Returns: "Done — meeting with Sarah scheduled for Tuesday at 2pm"

The user never sees any JSON or API calls. They speak naturally, and MCP handles the translation.

MCP vs. Other Integration Approaches

Aspect Zapier / n8n Custom APIs MCP
Complexity Low (no-code) High (custom dev) Medium (standard)
AI Awareness None — trigger/action Manual integration Native AI support
Context / Memory No Build it yourself Built-in
Multi-step Reasoning Limited branching Possible but complex Yes — agent-driven
Best For Simple automations Unique business logic AI agent ecosystems

When to Use Each

  • Low technical maturity (no APIs, no dev team) → Zapier / n8n
    Example: "Connect Gmail to a Slack notification when a VIP customer emails"

  • Medium technical maturity (some APIs, small dev team) → MCP + existing servers
    Example: "Let our AI assistant query Salesforce and schedule meetings via Google Calendar"

  • High technical maturity (full eng team, proprietary systems) → Custom APIs / MCP servers
    Example: "Build a custom MCP server connecting our AI to our proprietary ERP system"

Using the MCP Explorer Demo

The MCP Explorer in this repo (pages/4_MCP_Explorer.py) lets you:

  1. Walk through scenarios step by step — See exactly how each layer processes a request, with real JSON-RPC messages
  2. Compare approaches — Same task shown via Zapier, custom APIs, and MCP
  3. Understand the protocol — See the three-layer architecture and message format

No API key is needed — it's an educational simulation.

Running It

# If running the full app
streamlit run Home.py
# Then click "MCP Explorer" in the sidebar

# Or run it directly
streamlit run pages/4_MCP_Explorer.py

Assignment 3 Connection

The MCP Explorer directly supports Assignment 3, Question 3: How would your agent integrate with existing systems?

When writing your proposal, consider:

  • Which tools does your agent need to connect to? (Name them specifically.)
  • Which approach fits your organization? (Zapier, MCP, custom APIs — and why?)
  • What are the barriers? (Legacy systems, siloed data, no dev team?)
  • What's your phased rollout? (MVP with Zapier → scale with MCP?)

A strong answer names specific tools and justifies the integration approach based on your org's technical maturity.

Learn More

MIT Professional Education | Applied Generative AI for Digital Transformation