MCP Server

What Is MCP?

Core Concepts

As AI agents continue to evolve, developers increasingly face common challenges when building and integrating tools:

• No common standard, resulting in lower development efficiency.
• Repeated integrations across different platforms, increasing development time and maintenance costs.
• Inconsistent quality across community-built tools, making high-quality integrations difficult and expensive to maintain.

MCP (Model Context Protocol) was introduced to address these challenges. Developed by Anthropic, MCP is an open protocol that provides a standardized and secure way for large language models (LLMs) to communicate with external applications and data sources. It enables interoperable integrations and reduces fragmentation across AI ecosystems.

Think of MCP as a universal connector or USB interface for AI applications. It defines a standard protocol for exchanging data between applications and LLMs. Any AI model that supports MCP can interact with any MCP-compatible application.

Understanding Agents, MCP, and APIs (Tools) Through a Smart Assistant Analogy

Imagine having a smart personal assistant at home. You simply tell it what you need, and it takes care of everything for you. This is exactly the role of an AI Agent.

Role	In the Analogy	Technical Equivalent	Responsibility
👤 User	The homeowner	End user	Describes the task and waits for the result.
🤖 Agent	A smart personal assistant	AI agent	Understands the user's request, reasons through the task, and orchestrates the appropriate tools. • Responsibility: Determines what needs to be done and in what order. • Capability: Combines multiple tools into a complete workflow. • Think of it as: The brain and decision-making center.
🔌 MCP	A universal power outlet (like USB-C)	Protocol layer	Even the smartest assistant would struggle if every tool had a different interface. MCP standardizes how tools are discovered and invoked, allowing the agent to focus on solving tasks instead of adapting to different integrations. • Standardizes tool invocation • Hides implementation details from the agent • Enables plug-and-play interoperability
🛠️ Tool	Individual tools (Weather, Flights, Email, etc.)	API or service	Tools perform the actual work. • Weather tool → Returns weather information • Flight tool → Retrieves flight options • Email tool → Sends confirmation emails

Evolution of AI Integrations: API Documentation → Function Calling → MCP

Human Understanding → Model Understanding (Semi-Automated) → Direct Model Execution (Protocol-Based)

Aspect	API-based Integration	Function Calling	MCP-based Integration
🧠 Who Interprets It	Human developers	AI model + developer (manual tool registration)	AI model (automatic protocol discovery)
🎯 Primary Interface	Human-readable documentation	Function schema for AI models	Machine-readable protocol definition
🚀 How It Works	Read the API docs → Write requests → Call the API	Define function schemas → AI selects and calls functions	Expose an MCP server → AI automatically discovers and invokes capabilities
🧩 Tool Setup	Configure integrations separately for each platform	Register functions separately on each platform	Configure the MCP server once
🔍 Capability Discovery	Manual documentation lookup	Manual registration	Automatic capability discovery
🔄 Capability Updates	Documentation changes must be synchronized manually	Function definitions must be maintained on every platform	Capabilities are synchronized automatically through the MCP manifest
🧰 Integration Complexity	High — manual implementation and prompt engineering required	Medium — function schemas must be defined	Low — capabilities are exposed automatically through the protocol
🧠 Model Understanding	Prompt-driven, based on documentation	Infers usage from function descriptions	Structured protocol with built-in types for direct understanding
🔒 Invocation Safety	Low — incorrect parameters can easily cause failures	Medium — schemas validate parameter formats	High — protocol-level validation and contextual controls
💬 Analogy	"Read the manual and do it yourself."	"Read the function signature and call it."	"Plug in a USB-C cable and everything just works."
⚙️ Abstraction Layer	Application layer	SDK layer	Protocol layer
🌐 Extensibility	Low — platform-specific integrations	Medium — depends on platform support	High — unified capability discovery across platforms
👩‍💻 Developer Experience	Human-centric: write and maintain documentation	Semi-automated: register and maintain functions	Fully automated: integrate once, reuse across AI models

How to Use

APIfox MCP and HAP MCP serve different purposes depending on your integration needs:

• APIfox MCP allows large language models to read and understand the API documentation of HAP applications — ideal for “making the model API-aware.”
• HAP MCP enables direct access and manipulation of application data — designed for “executing business actions through the model.”

	APIfox MCP	HAP MCP
Purpose	Enable the model to understand HAP API documentation.	Allow the model to execute HAP application APIs.
Access	Documentation: https://apifox.mingdao.com	Accessible from within the Application API Documentation panel.
Use Case	When the goal is to understand available API structures for generating API call strategies or orchestrating workflows. 📍Example: For a Customer Case Management application, if you want to create a real-time data dashboard, you can let the model (via a client like Cursor) access the API documentation using APIfox MCP (preferably V3 APIs).	When the goal is to perform actions or access real-time data directly within a production environment. 📍Example: In an orchestration flow, the model may directly retrieve application data, perform analysis, map fields, or update records using HAP MCP.
Characteristics	Exposes API definitions, documentation, and examples (read-only). Helps the model act as a knowledgeable API assistant. Best suited for: project development, API discovery, documentation automation.	Exposes executable tools with permission and security control. Helps the model act as a capable operational agent. Best suited for: business execution, production system workflows, governance.

Configure MCP

Get the HAP MCP Endpoint

1. Open your application's API Documentation.
2. Switch to the MCP tab.
3. Copy the generated URL, which will be used in the following configuration steps.

Cursor

1. Open Cursor → Settings → Tools & MCP.
2. Click New MCP Server.

After saving, a green status indicator means the connection has been established successfully.

Dify

Use MCP in an Agent App

1. Click Plugins in the upper-right corner.
2. Explore Marketplace.
3. Search for MCP SSE / StreamableHTTP and install the plugin.

After installation, locate MCP SSE / StreamableHTTP in the installed plugins list.

Example configuration

Replace "hap-mcp" with any name you prefer, and replace the URL with the MCP endpoint generated by your HAP application.

{"hap-mcp":{"transport":"streamable_http","url":"<Generated MCP URL>"}}

Add the MCP tool and verify that it works as expected.

Use MCP in a Workflow

1. Click Plugins in the upper-right corner.
2. Explore Marketplace.
3. Search for Agent Strategies (Support MCP Tools) and install it.

Example configuration

Replace "hap-mcp" with any name you prefer, and replace the URL with the MCP endpoint generated by your HAP application.

{"hap-mcp":{"transport":"streamable_http","url":"<Generated MCP URL>"}}

VS Code + Cline

• Install the Cline Extension

  

• Click MCP Server to configure the MCP server

  

Example configuration

"hap-mcp": {
  "url": "<Generated MCP URL>",
  "type": "streamableHttp",
  "disabled": false,
  "autoApprove": []
}