--- title: "Why MCP exists: the integration problem it solves" description: "The N x M problem Before the Model Context Protocol, every AI application that wanted to use a tool or data source needed a custom integration. Five AI…" url: https://optimizeall.com/learn/model-context-protocol-mcp/why-mcp-the-integration-problem updated: 2026-10-05 --- Model Context Protocol (MCP): Connect AI to Your Tools and Data · Why MCP and how it works · lesson 1 of 18 · 13 min # Why MCP exists: the integration problem it solves ## The N x M problem Before the Model Context Protocol, every AI application that wanted to use a tool or data source needed a custom integration. Five AI apps and twenty business systems meant up to a hundred bespoke connectors, each with its own auth, schemas and bugs. Every new model or framework started from zero. **MCP** turns that into N + M. Each AI application implements one MCP **client**; each system is exposed once as an MCP **server**. Any compliant client can then use any compliant server. It is often compared to USB-C for AI: one standard port instead of a drawer of adapters. A closer analogy for engineers is the **Language Server Protocol**, which let every editor support every programming language through one protocol; MCP explicitly borrows ideas from it. ## A short history - **November 2024**: Anthropic open-sourced MCP with a specification and SDKs. - **2025**: rapid adoption across AI clients (Claude apps, IDEs and coding tools, ChatGPT, agent SDKs from multiple vendors) and thousands of community servers. The spec added Streamable HTTP, OAuth-based authorization, structured tool output, elicitation, and then (November 2025) tasks, URL-mode elicitation and an extensions framework. - **December 2025**: MCP was donated to the **Agentic AI Foundation (AAIF)**, a directed fund under the Linux Foundation, alongside Block's goose and OpenAI's AGENTS.md, giving it vendor-neutral governance. - **July 28, 2026**: the **2026-07-28** specification made the protocol core **stateless** (no initialize handshake, no protocol sessions), introduced Multi Round-Trip Requests, header-based routing and cacheable list results, hardened authorization, formalized extensions (Tasks, MCP Apps, enterprise authorization) and adopted a deprecation policy with a minimum twelve-month window. The Tier 1 SDKs (TypeScript, Python, Go, C#) support it. The specification is versioned by date. Clients and servers negotiate a version, and many deployed servers and clients still speak 2025-06-18 or 2025-11-25, so compatibility matters. This course teaches the durable concepts and flags where 2026-07-28 changed things. ## What MCP standardizes, and what it doesn't MCP standardizes **how context and capabilities are described and exchanged**: listing and calling tools, reading resources, fetching prompt templates, asking the user for input, transports, and authorization for remote servers. It does **not** decide which model you use, how your agent plans, or whether a tool call is wise. Those remain the host application's responsibility, including security decisions. ## Why businesses care - **Reuse**: build a CRM connector once; use it from Claude, an IDE, ChatGPT and your own agents. - **Governance**: a single layer where access, logging and approvals can be enforced. - **Speed**: teams wire new data sources into AI workflows in days instead of weeks. - **Ecosystem**: vendors increasingly ship official MCP servers for their products, and an official **MCP Registry** (in preview) catalogs publicly available servers. ## When MCP is not the answer - A single app calling a single internal API it owns can just use native function calling. - Ultra-low-latency inner loops may not want an extra protocol hop. - If you cannot secure a server properly (auth, least privilege, logging), don't expose it; MCP makes integration easy, including insecure integration. ## Worked example: a Karachi digital agency The agency uses Claude for strategy work, an AI-enabled IDE for its developers, and a custom reporting agent. It needs all three to access the same three systems: its CRM, a Google Sheets-based media plan, and an internal brand-guidelines library. - Without MCP: three apps × three systems = nine integrations to build and maintain. - With MCP: three servers (CRM, media plan, brand library), and each app connects as a client. Adding a fourth app costs zero new integrations. - Governance win: the CRM server exposes read-only tools to most users and a small set of write tools only to account leads, enforced in one place. ## Hands-on: see MCP in action in five minutes 1. Install the Python SDK with its CLI: `pip install "mcp[cli]"` (or `uv add "mcp[cli]"`). Note that `pip install mcp` now installs SDK v2; pin `mcp>=1.28,<2` if you must stay on v1. 2. Save this as `server.py`: ```python from mcp.server import MCPServer mcp = MCPServer("Demo") @mcp.tool() def add(a: int, b: int) -> int: """Add two numbers.""" return a + b @mcp.resource("greeting://{name}") def greeting(name: str) -> str: """Greet someone by name.""" return f"Hello, {name}!" if __name__ == "__main__": mcp.run() ``` 3. Run `uv run mcp dev server.py` (or `mcp dev server.py`) to open the **MCP Inspector**, call `add` with `a=1, b=2`, and read `greeting://Aisha`. Notice what you did not write: no JSON Schema (type hints are the schema), no request parsing and no protocol handling. ## Measuring success For an MCP program, track the number of integrations reused across clients, time to connect a new system, share of AI tool traffic flowing through governed servers, and incidents per server. ## Video lecture: Why MCP exists: the integration problem it solves Lecture coming soon · 14 chapters · about 9 minutes. Read the full transcript below. 1. Why MCP exists 2. Why it matters 3. N × M becomes N + M 4. A short history 5. The 2026-07-28 spec 6. Simple example: one issue tracker, two AI apps 7. What MCP standardizes 8. Business value and limits 9. Example: Karachi agency 10. Hands-on in five minutes 11. Deeper: the Karachi agency (illustrative) 12. Watch me do it: the demo server 13. Try this now 14. Recap ## Lecture transcript ### Why MCP exists Imagine a drawer full of chargers, one for every device you've ever owned. That's what AI integrations looked like before the Model Context Protocol. Every AI app needed its own custom connector for every tool and data source. In this lesson you'll learn what problem MCP solves, where it came from, what it does and doesn't standardize, and when you shouldn't use it. ### Why it matters Why should you care? Because integrations are where AI projects quietly stall. The model works in the demo, then someone asks, can it see our CRM, and the answer is a six week custom project. Multiply that by every tool and every AI app, and you understand why so many pilots never scale. MCP's promise is simple: build each connection once, and every compatible AI application can use it. That changes the economics of AI adoption for agencies, SaaS vendors and enterprises alike. ### N × M becomes N + M Here's the math. Five AI applications and twenty business systems could mean up to a hundred custom integrations, each with its own authentication, data formats and bugs. MCP changes the equation. Each AI application implements one MCP client. Each system is exposed once as an MCP server. Any compliant client can use any compliant server, so five plus twenty is twenty five pieces, not a hundred. Engineers often compare it to the Language Server Protocol, which let every code editor support every programming language through one protocol. ### A short history A quick history. Anthropic open sourced MCP in November twenty twenty four. Through twenty twenty five it spread across AI clients, from Claude's apps and coding tools to ChatGPT and agent SDKs from several vendors, and thousands of community servers appeared. The spec added Streamable HTTP, OAuth based authorization, structured tool output and elicitation. In December twenty twenty five, MCP was donated to the Agentic AI Foundation under the Linux Foundation, giving it neutral governance. Then on July twenty eighth, twenty twenty six, a major new version made the protocol core stateless. ### The 2026-07-28 spec That July twenty twenty six release is worth a closer look. It removed the initial handshake and protocol level sessions, so every request describes itself and can land on any server behind a load balancer. It introduced multi round trip requests for asking users for input, header based routing, cacheable tool lists, stronger authorization, a formal extensions framework and a twelve month deprecation policy. The official TypeScript, Python, Go and C sharp SDKs support it. But many servers and clients still speak earlier versions, so compatibility matters, and this course flags the differences as we go. ### Simple example: one issue tracker, two AI apps Here's a simple example. You use two AI tools: Claude for writing and an AI coding assistant in your editor. Both need to read your project's issue tracker. Before MCP, each tool needed its own plugin, built by someone, maintained by someone, and often missing features. With MCP, the issue tracker vendor publishes one MCP server. You add it to Claude and to your editor, and both can search issues, read details and draft updates, through exactly the same tools. When the vendor adds a feature, both apps get it. ### What MCP standardizes Be clear about scope. MCP standardizes how context and capabilities are described and exchanged: listing and calling tools, reading resources, fetching prompt templates, asking the user for input, transports and authorization for remote servers. It doesn't choose your model, plan your agent's steps, or decide whether a tool call is wise. Those remain the host application's job, including security decisions. MCP makes integration easy, and that includes insecure integration, so governance is on you. ### Business value and limits Why do businesses care? Reuse: build a CRM connector once and use it from Claude, an IDE, ChatGPT and your own agents. Governance: one layer where access, logging and approvals can be enforced. Speed: new data sources wired into AI workflows in days. And ecosystem: vendors increasingly ship official MCP servers, and an official MCP Registry, currently in preview, catalogs publicly available servers. But skip MCP when a single app calls a single API it owns, when you need ultra low latency inner loops, or when you can't secure the server properly. ### Example: Karachi agency A worked example. A digital agency in Karachi uses Claude for strategy, an AI enabled IDE for developers, and a custom reporting agent. All three need the CRM, a media plan in Google Sheets, and a brand guidelines library. Without MCP, that's nine integrations. With MCP, it's three servers, and each app connects as a client. A fourth app costs nothing extra. And governance improves: the CRM server gives read only tools to most people and write tools only to account leads, enforced in one place. ### Hands-on in five minutes Now try it yourself. The lesson has a twelve line Python server with one tool that adds two numbers and one resource that greets someone by name. Install the Python SDK with its command line extra, save the file, and run mcp dev to open the MCP Inspector. Call the add tool and read a greeting. Notice what you didn't write: no JSON schema, because type hints become the schema, no request parsing and no protocol code. The SDK does all of it. ### Deeper: the Karachi agency (illustrative) Let's deepen the Karachi agency example with illustrative numbers. Before MCP, the agency's developers spent about two weeks building each custom integration, and they had built five, some already broken by vendor API changes. With MCP, they built three servers in roughly the time of one old integration, because the SDK handled the protocol and every client reused them. The bigger surprise was governance. For the first time, the operations lead could answer, which AI tools can touch client data, by reading one catalog of three servers and their scopes. When a new AI writing tool arrived that spoke MCP, connecting it took an afternoon rather than a sprint. ### Watch me do it: the demo server Watch me do it. Let's walk through the twelve line demo server. First line imports MCP server from the SDK. Then I create the server and name it Demo. The add function has a tool decorator, two integer parameters and a docstring. That's all the SDK needs: the type hints become the input schema and the docstring becomes the description. The greeting function has a resource decorator with a URI template, greeting colon slash slash name, so any name becomes a readable resource. At the bottom, under the main guard, mcp dot run starts it over standard I O. Now I run mcp dev on the file. The Inspector opens, lists one tool and one resource template. I call add with one and two, and get three back, with structured content. Then I read greeting slash Aisha and get, Hello, Aisha. Five minutes from nothing to a working, standards based integration. ### Try this now Try this now. Make two lists on paper. On the left, every AI application your team uses or plans to use. On the right, every business system people wish those apps could reach: CRM, analytics, docs, ticketing, finance. Count the lines you'd need without MCP, and the boxes you'd need with it. Then circle two systems on the right that would create the most value if every AI app could use them. Those are your first MCP servers, and the rest of this course shows you how to build them properly. ### Recap To recap: MCP turns N times M integrations into N plus M, standardizes the exchange of tools, resources and prompts, is governed by a neutral foundation, and just became stateless with the twenty twenty six spec. Your next step: list the AI apps and business systems in your organization, compare N times M with N plus M, and choose the first two systems you'd expose as MCP servers. ## Key takeaways - MCP turns N x M custom integrations into N clients plus M servers. - It standardizes how tools, resources and prompts are described and exchanged, not how agents reason. - MCP is governed by the Agentic AI Foundation under the Linux Foundation. - The 2026-07-28 spec made the core stateless and formalized extensions; older versions are still widely deployed. - Skip MCP for single-app, single-API cases or when you cannot secure the server. ## Try it List the AI apps and business systems in your organization, calculate N x M versus N + M, and pick the first two systems you would expose as MCP servers. - [Next: MCP architecture: hosts, clients, servers and the protocol](https://optimizeall.com/learn/model-context-protocol-mcp/mcp-architecture) - [All lessons of Model Context Protocol (MCP): Connect AI to Your Tools and Data](https://optimizeall.com/learn/model-context-protocol-mcp)