In my recent posts on tool calling, we explored how AI agents decide which tool to use and how to use it. We saw the model generate a tool call, our code execute it, and the result feed back to the model. This works well enough for tutorials, but it glosses over a critical question: Where do the tools come from?
In every example so far, we defined the tools ourselves, by hand, in the same Python script as the agent. For a tutorial, that's fine, but in a real application, this approach quickly falls apart. Every tool requires a custom integration, and every new model-tool pair demands its own connector. Consider a setup with three AI models and ten tools—that’s potentially thirty integrations to maintain. Every time a model or tool changes, something could break. In other words, as neat as this approach is for a simple demo, it doesn’t scale at all. 🤷♀️
This isn’t a niche problem—it’s a major challenge of the agentic AI era. And it’s exactly what the Model Context Protocol (MCP) was designed to solve.
What About MCP?
Before MCP, connecting an AI model to an external tool (like a database, a file system, Slack, or GitHub) meant writing a custom integration every single time. The model had to know the tool’s specific calling format, and the tool had to respond in a format the model could understand. If the model changed, you rewrote the integration. If a new tool appeared, you started from scratch.
This is the M×N problem: with M models and N tools, you need M×N custom integrations. As the ecosystem grows, this becomes unsustainable.
!M×N integration problem diagram
A helpful metaphor is the evolution of computer hardware. In the early days, every peripheral—printer, mouse, keyboard—had a proprietary connector. A printer bought for one computer wouldn’t work with another. Then USB arrived as a single, standard connector: any device that supports USB works with any computer that supports it.
Think of MCP as the “USB for AI agents.” One standard protocol, and any agent that supports it can connect to any tool that supports it—regardless of vendor. As of 2026, MCP has become widely adopted across major AI platforms, cloud services, and open-source frameworks, much like USB supplanted proprietary connectors.
The Model Context Protocol is an open standard that enables developers to build secure, two-way connections between data sources and AI-powered tools. Created by Anthropic and open-sourced in November 2024, it replaces point-to-point integrations with a single client-server protocol. Any MCP-compatible AI host can discover and use any MCP-compatible tool or data resource.
More precisely, the MCP architecture has three main participants: