The Model Context Protocol (MCP) is the open protocol that standardizes how large language models (LLMs) connect to external data sources and tools. Released by Anthropic in November 2024 under the MIT license, MCP evolved in roughly a year and a half from an experimental specification into the AI industry's de facto standard for model-tool integrations. This lesson explores the problem MCP solves, its history, its adoption dynamics, and its fundamental impact on how we build AI systems.
The M×N Problem: Why We Needed MCP
The pre-MCP integration explosion
Before MCP, every AI model provider implemented its own mechanism for interacting with external tools. OpenAI had Function Calling (launched in June 2023), Anthropic had Tool Use, Google had Function Declarations in Gemini, and every framework — LangChain, LlamaIndex, CrewAI — had its own abstractions. The result was a fragmented ecosystem:
The pre-MCP situation (2023-2024):
Claude ──→ Custom tool handler for PostgreSQL
Claude ──→ Custom tool handler for GitHub
Claude ──→ Custom tool handler for Slack
GPT ────→ Custom function calling for PostgreSQL
GPT ────→ Custom function calling for GitHub
GPT ────→ Custom function calling for Slack
Gemini ──→ Custom function declaration for PostgreSQL
Gemini ──→ Custom function declaration for GitHub
Gemini ──→ Custom function declaration for Slack
Total: 3 models × 3 tools = 9 unique integrations
M models × N tools = M×N integrations
Every integration requires custom code, separate documentation, dedicated tests, and independent maintenance. At enterprise scale — with 5-10 AI models and 20-50 internal tools — you ended up with hundreds of point-to-point integrations, each with its own bugs and limitations.
The real cost of fragmentation
Fragmentation was not just an engineering problem — it was a business problem. The typical pre-MCP scenario for an enterprise team looked like this:
- Months of development to integrate a few AI models with internal systems
- Engineers dedicated exclusively to integration maintenance — every API update from a provider triggers cascading changes
- Severe vendor lock-in: switching from one model to another requires rewriting most of the integration code
- Zero portability: tools built for one model did not work with another
The problem was identical to the one in the hardware industry before USB-C: every manufacturer had its own connector, and consumers (in our case, developers) paid the price of fragmentation.
The USB-C analogy — and its limits
The "USB-C for AI" metaphor is useful but imperfect. USB-C standardizes a physical connector — a single plug type for power, data, and video. MCP standardizes a communication protocol — a set of rules through which AI models discover, invoke, and receive results from external tools.
The analogy works along three dimensions:
| Dimension | USB-C | MCP |
|---|---|---|
| Universality | One connector → any device | One protocol → any model + tool |
| Complexity reduction | Eliminates the drawer full of cables | Eliminates custom code per integration |
| Adoption through utility | Adopted because it is better | Adopted because it is more efficient |
The analogy breaks down when we talk about complexity. USB-C is "plug and play" — you connect the cable and it works. MCP requires implementing servers, configuring transports, managing authentication, and validating security. MCP is closer to being the AI equivalent of the HTTP protocol — a communication standard on top of which complex systems are built, not a connector you plug in and forget.
The MCP Solution: M + N Instead of M × N
The fundamental principle
MCP solves the M×N problem by introducing a standardized abstraction layer:
The MCP solution:
Claude ──→ MCP Client ─┐
GPT ─────→ MCP Client ──┼──── MCP Protocol ────┬── MCP Server PostgreSQL
Gemini ──→ MCP Client ─┘ ├── MCP Server GitHub
Llama ───→ MCP Client ──────────────────────────├── MCP Server Slack
└── MCP Server Jira
Total: 4 clients + 4 servers = 8 implementations (instead of 16)
M clients + N servers = M + N (instead of M × N)