格 千格 KiloCell

MCP Complete Guide to Agreements

Model Context Protocol: "USB-C" in the AI field

MCP Communications Architecture Map (interactive)

MCP Client

Claude / Cursor

⟷

JSON-RPC

Stdio / SSE

⟷

MCP Server

Files / DB / API

📄 Resources
🔧 Tools
💬 Prompts

Resources

This is AI. (c) static or dynamic data that can be read. For example, the content of a log file or a snapshot of a database table. Resources are read-only, like... AI's Sensory, which allows it to observe the outside world.

🔧 Tools (tools)

This is an action that AI can execute. Tools Allow AI Impacts on the outside world, such as " Create files", "Send mail" or "Enforce SQL updates". Every tool has a strict JSON. Schema defines its input parameters.

💬 Prompts (Phrasing Template)

Allows the server to provide predefined instructions to the client. This is very useful in building helpers in specific areas, such as the code audit specialist hint template, which automatically guides AI How to use the relevant audit tools.

MCP protocol depth resolution: How did the AI-era “one-size-fits-all interface” come about?

In AI 2.0 The age, the large model (LLM), is no longer just a probabilistic model for predicting the next word, is evolving into a “brain” capable of operating the real world. And... Model Context Protocol (MCP) And that's the standard neurofibre that connects the brain to the outside world. MCP by Anthropic Agreement, which aims to eliminate AI by a standardized, safe and easily extended approach Information gap between models and external data sources. This paper will be developed from the bottom of the technology architecture to the development model, and will be used to provide a deep insight into the MCP. The core logic.

Why do we need MCP? Say goodbye to the "glap code" age.

Before MCP appears, if you want AI to access your database, you must have every AI client (e. g. different IDE or Web) (a) To develop a special “glubber code” and an API adapter. This “one-on-one” docking pattern is extremely inefficient. The core contribution of MCP is that it defines a “multi-to-multi” standard: Any MCP server can be connected to any client that supports MCP. Like USB-C, which unites the physical interfaces of electronics, MCP. Harmonizes the context access protocol for AI.

2. Technical bottom: JSON-RPC 2.0 and transmission mechanisms

MCP Agreement based on mature JSON-RPC 2.0 Normative. This lightweight agreement based on request-response and notification is very well suited to AI, which is a highly reliant, insular interaction.

  • Transfer layer flexibility: Support stdio (locally run) and SSE (Server-Sent Events, remote connection).
  • Initial handshake: Client and server do it when connection is established initialize Exchange, dynamically negotiate each other's “capabilities”.

3. Three pillars: Resources, Tools and Prompts

The key to understanding MCP is to understand the three core concepts it defines:

  1. Resources: AI's Sensory, read-only data stream.
  2. Tools (tools)： AI's Hands, an enforceable move.
  3. Prompts (Phrasing Template): AI's "Set of Instructions", predefined interactive templates.

4. Field Vision: How to design a high-quality MCP Tool?

Operational recommendations: addressing the problem of "AI blind attempts"

  • Specific issues: AI often fails when calling the tool because the parameters are not regulated or misinterpreted.
  • Optimization of programmes: Detailed description (Description) is essential when designing MCP Tool. Using JSON Schema required and enum Containment, intercept invalid requests at the level of the agreement, significantly increase the success rate of AI calls.

5. Development and deployment: TypeScript and Python SDK

Anthropic officially provides a perfect SDK that significantly lowers the development threshold:

// TypeScript SDK Example
import { Server } from "@modelcontextprotocol/sdk/server";
const server = new Server({ name: "my-server", version: "1.0.0" }, { capabilities: { tools: {} } });

6. Summary: future prospects for the MCP agreement

The emergence of MCP marked the evolution of AI ecology from “single islandization” to “standardization”. I'm sure I'm going to be able to get the MCP agreement in place for the next 10 years. Application development core paradigm. Thousands will continue to interpret the MCP protocol evolution and best practices for you.