What is MCP?
What is MCP?
Section titled “What is MCP?”The Model Context Protocol (MCP) is a standardized interface and framework that enables AI models to seamlessly connect with external tools, resources, and environments. Published by Anthropic as an open standard in November 2024, MCP allows AI applications to communicate with diverse data sources and tools in a consistent way.
Analogies for Understanding MCP
Section titled “Analogies for Understanding MCP”The Interpreter Analogy
Section titled “The Interpreter Analogy”Imagine someone who only speaks English trying to have a conversation with people who speak French, German, and Spanish. Direct communication is impossible because they speak different languages. What they need is an interpreter.
AI models face a similar situation. Large Language Models (LLMs) can only answer based on the knowledge contained in their training data. They cannot directly access external information or tools such as real-time weather, breaking news, or corporate databases.
MCP plays the role of that interpreter. It stands between AI agents and various external tools and resources, enabling bidirectional communication through a common language — the protocol.
The USB-C Analogy
Section titled “The USB-C Analogy”I think of MCP as USB-C for AI systems.
Before USB-C, connecting devices required different cables and standards for each piece of hardware. With the widespread adoption of USB-C as a common standard, a single cable can now connect smartphones, laptops, displays, and many other devices.
MCP provides the same kind of universal standard. AI applications connect to diverse external tools — databases, file systems, web APIs, code execution environments — through a single “port” called MCP.
The Problems MCP Solves
Section titled “The Problems MCP Solves”LLMs excel at generating text, but they have the following limitations:
| Limitation | Description |
|---|---|
| Knowledge cutoff | No awareness of events that occurred after training completed |
| Real-time information | Cannot access current weather, stock prices, or news |
| External system integration | Cannot directly operate databases or APIs |
| State persistence | Cannot make permanent changes such as writing to files |
MCP supplements these limitations, enabling AI agents to work with real-world tools and information.
The Three Components of MCP
Section titled “The Three Components of MCP”MCP consists of three components. Each is explained in detail on its own page.
| Component | Role | Details |
|---|---|---|
| Host | User-facing AI application | Architecture |
| Client | Component handling communication with MCP servers | Architecture |
| Server | External program that provides tools and resources | Architecture |
MCP servers can also provide three types of capabilities:
| Capability | Overview | Details |
|---|---|---|
| Tools | Functions an AI can invoke (with effects/side effects) | Capabilities |
| Resources | Read-only data provision | Capabilities |
| Prompts | Predefined prompt templates | Capabilities |
Summary
Section titled “Summary”- MCP is a standardized protocol connecting AI models to external tools and resources
- Like an interpreter or USB-C, it provides a common connection standard between different systems
- It solves the challenges of LLM knowledge limitations, real-time information access, and external system integration
Next Steps
Section titled “Next Steps”- Why MCP? — The M×N integration problem and how MCP solves it
- MCP Architecture — Roles and structure of Host, Client, and Server with diagrams
- MCP Capabilities — Differences and use cases for Tools, Resources, and Prompts
Frequently Asked Questions
Section titled “Frequently Asked Questions”Q: Which company developed MCP?
A: Anthropic developed MCP and published it as open source in November 2024. It is now an open standard, with various companies and communities building compatible tools.
Q: What do I need to use MCP?
A: I need an AI application that implements an MCP client (such as Claude Desktop or Cursor) and an MCP server. When using existing MCP servers, I can connect from an AI app without additional programming.
Q: How is MCP different from an API (Application Programming Interface)?
A: An API defines a connection interface to a specific service. MCP is a meta-protocol that standardizes communication between AI models and tools — MCP servers often call APIs internally. MCP does not replace APIs; rather, it makes it easier for AI to use APIs.
References
Section titled “References”Link to this page (Japanese): MCPとは