What is MCP? A standard connecting AI to external tools like USB
MCP (Model Context Protocol) is an open standard protocol established by Anthropic in November 2024. Its objective is to standardize how AI models connect to external tools, data sources, and systems.
A popular metaphor is "USB-C for AI." Just as USB allows any brand of computer to connect to devices using the same physical cable, MCP provides a unified mechanism for any AI model to invoke external tools using the exact same procedure.
Before MCP arrived, integrating AI with external systems required rewriting implementations to fit each AI provider's proprietary specifications. Even when having Claude and GPT-4 query the identical database, separate integration code was required for each. MCP eliminates this inefficiency. The core value proposition of MCP is delivering a world where "you implement once, and it works across any MCP-compatible AI."
Rapid industry adoption
Following its adoption by OpenAI in April 2025, MCP rapidly established itself as the industry standard.
| Timeframe | Key milestones |
|---|---|
| November 2024 | Anthropic announces MCP and open-sources it |
| April 2025 | OpenAI officially adopts MCP |
| July 2025 | Microsoft integrates MCP into Copilot Studio |
| November 2025 | AWS Bedrock adds support for MCP |
| December 2025 | Anthropic donates MCP to the AAIF under the Linux Foundation |
| March 2026 | Monthly SDK installations exceed 97 million, with over 5,800 community-built servers |
As of April 2026, all major AI providers support MCP, making it the de facto industry standard.
MCP architecture: three key actors and two layers
MCP is architected around a straightforward client-server model featuring three primary actors.
Roles of Host, Client, and Server
┌──────────────────────────────────────────────┐
│ MCP Host（AIアプリケーション） │
│ 例: Claude Desktop, Claude Code, VS Code │
│ │
│ ┌─────────────┐ ┌─────────────┐ │
│ │ MCP Client 1│ │ MCP Client 2│ ... │
│ └──────┬──────┘ └──────┬──────┘ │
└─────────┼─────────────────┼──────────────────┘
│専用接続 │専用接続
┌─────▼──────┐ ┌──────▼──────┐
│ MCP Server │ │ MCP Server │
│（ローカル） │ │（リモート） │
│例:ファイルシ│ │例: Sentry, │
│ステム,DB │ │ GitHub API │
└─────────────┘ └─────────────┘An MCP Host is an AI-powered application such as Claude Desktop or VS Code. It spawns an MCP Client for each MCP server it connects to, with each client maintaining a dedicated one-to-one connection. An MCP Server is a program providing access to external tools or data sources, and can reside either locally or remotely.
Data layer and transport layer
MCP is structured across two layers.
The data layer defines the messaging protocol based on JSON-RPC 2.0. This layer defines the "primitives" that form the core concept of MCP.
| Primitive | Role | Examples |
|---|---|---|
| Tools | Functions the AI can execute | File operations, API calls, database queries |
| Resources | Data referenced as context | File contents, database schemas, logs |
| Prompts | Reusable templates | System prompts, few-shot examples |
The transport layer manages communication channels. MCP supports two types of transports.
- Stdio (Standard Input/Output): Local inter-process communication on the same machine. Extremely fast with zero network overhead
- Streamable HTTP: Remote communication combining HTTP with Server-Sent Events. Supports authentication mechanisms such as OAuth and API keys
Connection lifecycle
The workflow for an AI to utilize an MCP server proceeds as follows:
- Initialization: The Client sends an
initializerequest to the Server to negotiate protocol versions and supported capabilities - Tool discovery: The Client retrieves the list of tools provided by the Server using
tools/list - Tool execution: When the AI decides to invoke a tool, the Client dispatches a request to the Server via
tools/call - Notification: If server-side tools change, the Client is notified in real time via
notifications/tools/list_changed
Thanks to this architecture, AI models can dynamically discover server capabilities and consume newly added tools without requiring application redeployment.
Differences from Function Calling: which should you use?
A concept frequently confused with MCP is "Function Calling." Here is an overview of how the two relate.
What is Function Calling?
Function Calling is a native capability built into LLMs themselves. The model parses user intent, evaluates whether a predefined function should be called, and outputs structured JSON. The actual function execution is carried out by the application layer.
Characteristics:
- Tool definitions must be transmitted with every LLM request
- Implementations are tightly coupled to specific AI providers
- Fast to implement for simple use cases
Comparison between MCP and Function Calling
| Dimension | Function Calling | MCP |
|---|---|---|
| Scope | Feature of an individual model | Cross-provider industry standard |
| Tool definition | Transmitted per request | Dynamically supplied by the server |
| Portability | Provider-dependent | Implement once, reuse everywhere |
| Architecture | Integrated within the LLM | Independent server process |
| Best suited for | Small scale, single-model apps | Large scale, multi-model systems |
Selection guidance
When Function Calling is preferable: Small projects contained within a single AI provider, with established APIs, limited tool counts, or rapid prototype development.
When MCP is preferable: Systems utilizing multiple AI models, setups with large numbers of independently updated tools, applications requiring flexibility to switch AI providers, or building enterprise-grade AI agent foundations.
MCP is not a drop-in replacement for Function Calling; rather, it is the optimal choice when prioritizing scalability and portability. Internally, it often relies on Function Calling mechanics while acting as an abstraction layer independent of any single provider.
Practical MCP server use cases
As of April 2026, over 5,800 official and community-developed MCP servers have been released. Here are some representative use cases.
Development tool integration
With the GitHub MCP Server, AI can directly create repositories, perform code reviews, manage issues, and open pull requests. Connecting it to Claude Code or GitHub Copilot allows AI to seamlessly navigate across your entire development workflow.
The Sentry MCP Server connects to error tracking systems. With just a natural language prompt like "Fix the error that occurred most frequently in the last 24 hours," the AI can search Sentry for the error, pinpoint the relevant code, and suggest a fix.
Business tool integration
Using the Google Drive / Docs MCP Server, AI can search and reference internal documents while generating responses. Without having to build your own RAG pipeline from scratch, you can use your existing Google Drive as an AI knowledge base.
With the Slack MCP Server, AI can reference channel conversation histories or post messages, functioning as an active participant in Slack.
Database and analytics integration
With PostgreSQL / MySQL MCP Servers, AI can directly run SQL queries to retrieve and analyze data. It enables use cases where an instruction like "Find out why last month's sales dropped compared to the previous month" leads the AI to construct queries and generate a report.
With the official AWS MCP Server collection released by AWS, AI can operate AWS services such as S3, Lambda, and CloudWatch. This lets you delegate command-line infrastructure management tasks to AI.
MCP as an AI agent foundation
The reason MCP means more than just "standardizing tool calls" is that it was designed as the foundational infrastructure for AI agents.
Why agents need MCP
AI agents do not just answer one-off questions; they autonomously execute multi-step tasks. To do that, they must dynamically understand "what tools are available" and select the right tool depending on the situation.
MCP's mechanisms for "tool discovery (tools/list)" and "real-time notifications" were designed precisely to meet this need. An agent can expand its capabilities without downtime simply by connecting a new MCP server at runtime.
As discussed in The Argument Against Low-Code in the Age of AI Agents, in an era where AI can directly operate code and systems, the abstraction layers of no-code tools can become "constraints." MCP embodies the opposite philosophy. It enables controlled openness by giving AI access rights to systems through explicit interfaces (MCP Servers).
The multi-agent context
In multi-agent systems where multiple AI agents work in coordination, sharing information and dividing roles among agents pose challenges. MCP provides a standardized toolset that agents can share, dramatically simplifying multi-agent architecture design.
Even in practical examples like Automating Project Management with Claude Code, the design overhead of connecting AI to multiple tools was a key hurdle, but the standardization of MCP is systematically resolving this issue.
MCP's current status and outlook
Enterprise readiness in 2026
In December 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF) under the Linux Foundation. Because a neutral organization co-founded by Anthropic, OpenAI, and Block governs the specification, the risk of vendor lock-in has been reduced.
Enterprise-grade enhancements are also progressing. The Q2 2026 roadmap includes OAuth 2.1 support and integration with enterprise identity providers such as Okta and Azure AD. As authentication and authorization standards advance, companies will find it easier to meet security requirements when connecting AI to internal systems via MCP.
How developers can start right now
If you have an MCP-compatible AI application (such as Claude Desktop or VS Code + GitHub Copilot), you can try out publicly available MCP servers immediately.
To make your internal tools MCP-compatible, implement an MCP server using the SDKs released by Anthropic (for Python, TypeScript, Java, C#, and more). Simply wrapping your existing APIs or data sources with an "MCP wrapper" makes them accessible to any MCP-compatible AI.
Basic steps for building a custom MCP server:
- Install the SDK (e.g.,
pip install mcp) - Define tools, resources, and prompts
- Select a transport (Stdio for local, HTTP for remote)
- Use the official MCP Inspector for testing
Conclusion
MCP is the industry standard for allowing AI to interact with any external tool in a unified manner. Surpassing 97 million installs in the year and a half since its November 2024 announcement, and adopted by OpenAI, Google, Microsoft, and AWS, it is no longer merely one option, but has become an underlying premise of AI system design.
A significant gap in the speed and quality of AI agent design will emerge between developers who understand and leverage MCP and those who do not.
- Unified interface: An MCP server implemented once can be used by any AI
- Dynamic tool discovery: Agents autonomously understand available tools at runtime
- Provider-agnostic: Avoids lock-in to specific AI vendors
- Ecosystem: Over 5,800 public servers ready for immediate use
When introducing AI into your business in earnest, MCP is the layer that fundamentally transforms how external tool connections are designed. Making internal systems MCP-compatible marks the first step in system design for the AI era.
Please feel free to contact us regarding AI-to-external-tool integration design or building an AI agent foundation powered by MCP.
