Research Notes

What is Model Context Protocol (MCP)?

A clear breakdown of Model Context Protocol — what it is, how it works, and why it's turning chatbots into real AI agents.

Published 19 March 2026Updated 20 October 20262 min read

If you've been following AI developments lately, you've probably heard the term MCP floating around. Here's a clean breakdown of what it actually means.

🔤 Breaking Down the Name

Model → the AI model (e.g., Claude, GPT)
Context → the information and tools the model has access to during a conversation
Protocol → a standardized ruleset for how two systems communicate

Note: In plain English: MCP is the standard way to plug things into an AI's brain.

🔌 The USB Analogy

Think of MCP like USB for AI. Instead of every application writing custom integrations with every AI model, MCP provides a universal connector:

flowchart LR
    Client["🤖 MCP Client\n(Claude Desktop, Cursor...)"]:::client
    Server1["🗂️ File System\nMCP Server"]:::server
    Server2["🐙 GitHub\nMCP Server"]:::server
    Server3["📅 Calendar\nMCP Server"]:::server
    Server4["🗄️ Database\nMCP Server"]:::server

    Client <--> Server1
    Client <--> Server2
    Client <--> Server3
    Client <--> Server4

    classDef client fill:#4A90D9,stroke:#2c5f8a,color:#fff
    classDef server fill:#5BA85A,stroke:#3a6e39,color:#fff

MCP Servers expose tools and resources (file systems, databases, APIs, GitHub, Slack...)
MCP Clients are AI applications (Claude Desktop, Cursor, etc.) that connect to those servers
The AI can then call those tools directly during a conversation

🚀 Key Applications

🗂️ Data & Files

Read/write local files, databases, cloud storage
Query SQL databases in natural language
Search through documents and codebases

💻 Developer Tools

GitHub/GitLab — read issues, create PRs, check CI status
IDE integration — AI reads your entire project context
Run terminal commands and execute code

📅 Productivity

Google Workspace — emails, calendar, Docs
Notion, Obsidian, Linear, Jira
Slack/Discord — read channels, send messages

🤖 Agent Workflows

Chain multiple tools together for autonomous task completion
Example: read email → check calendar → draft reply → send — all in one go

🤖 MCP vs. AI Agents — Not the Same Thing

A common misconception: MCP is not the principle behind AI agents. They're related, but distinct.

	What it is
AI Agent	The concept — AI that autonomously plans and takes actions
MCP	A tool — standardized way for agents to access the world

What actually makes an AI an agent:

LLM — the reasoning brain
Tools/Actions — ability to do things (MCP helps here)
Memory — context across steps
Loop — observe → think → act → repeat

Warning: MCP solves piece #2 (Tools), but the broader agent architecture involves all four. An agent without memory or a feedback loop is just a chatbot with buttons.

A good analogy: an agent is like a person, and MCP is like hands and a phone — the means to interact with the world.

💡 Why It Matters

MCP is what turns a chatbot into an actual agent capable of taking real-world actions. Before MCP, every AI tool integration was bespoke and brittle. Now, build a tool once as an MCP server and any MCP-compatible AI can use it.

Tip: It's a small protocol with a big implication: AI stops being a Q&A machine and starts being a genuine collaborator.

Part of my AI tooling series. Next: building your own MCP server in under 50 lines of Python.