MCP - Introduction
Background
- Generative AI applications are a great step forward as they often let the user interact with the app using natural language prompts.
- However, as more time and resources are invested in such apps, you want to make sure you can easily integrate functionalities and resources in such a way that it’s easy to extend, that your app can cater to more than one model being used, and handle various model intricacies.
- In short, building Gen AI apps is easy to begin with, but as they grow and become more complex, you need to start defining an architecture and will likely need to rely on a standard to ensure your apps are built in a consistent way.
- This is where MCP comes in to organize things and provide a standard.
What Is the Model Context Protocol (MCP)?
- The Model Context Protocol (MCP) is an open, standardized interface that allows Large Language Models (LLMs) to interact seamlessly with external tools, APIs, and data sources.
- It provides a consistent architecture to enhance AI model functionality beyond their training data, enabling smarter, scalable, and more responsive AI systems.
Why Standardization in AI Matters
- As generative AI applications become more complex, it’s essential to adopt standards that ensure scalability, extensibility, maintainability, and avoiding vendor lock-in. MCP addresses these needs by:
- Unifying model-tool integrations
- Reducing brittle, one-off custom solutions
- Allowing multiple models from different vendors to coexist within one ecosystem
- Note: While MCP bills itself as an open standard
- there are no plans to standardize MCP through any existing standards bodies such as IEEE, IETF, W3C, ISO, or any other standards body.
Learning Objectives
- Define Model Context Protocol (MCP) and its use cases
- Understand how MCP standardizes model-to-tool communication
- Identify the core components of MCP architecture
- Explore real-world applications of MCP in enterprise and development contexts
Why the Model Context Protocol (MCP) Is a Game-Changer
MCP Solves Fragmentation in AI Interactions
Before MCP, integrating models with tools required:
- Custom code per tool-model pair
- Non-standard APIs for each vendor
- Frequent breaks due to updates
- Poor scalability with more tools
Benefits of MCP Standardization
| Benefit | Description |
|---|---|
| Interoperability | LLMs work seamlessly with tools across different vendors |
| Consistency | Uniform behavior across platforms and tools |
| Reusability | Tools built once can be used across projects and systems |
| Accelerated Development | Reduce dev time by using standardized, plug-and-play interfaces |
High-Level MCP Architecture Overview
MCP follows a client-server model, where:
| Key | Desc |
|---|---|
| MCP Hosts | run the AI models |
| MCP Clients | initiate requests |
| MCP Servers | serve context, tools, and capabilities |
Host
- 职责：承载和执行 LLM 推理引擎，管理模型访问与凭证
- 控制域：模型选择、权限策略、用户授权流、会话协调
- 典型实例：Claude Desktop、Cursor IDE、自定义 AI 应用
Client
- 职责：代表 Host 发起与 Server 的连接，执行协议握手与能力协商
- 暴露能力：Sampling（LLM 生成）、Roots（文件系统边界）
- 通信模式：1:1 映射到单个 Server，通过 stdio 或 SSE 传输
Server
- 职责：暴露外部系统/数据源的结构化接口
- 核心原语：Resources（只读上下文）、Tools（可执行函数）、Prompts（可复用指令模板）
- 设计模式：单一职责原则，每个 Server 封装一个外部系统
Key Components
| Components | Desc |
|---|---|
| Resources | Static or dynamic data for models |
| Prompts | Predefined workflows for guided generation |
| Tools | Executable functions like search, calculations |
| Sampling | Agentic behavior via recursive interactions |
| Elicitation | Server-initiated requests for user input |
| Roots | Filesystem boundaries for server access control |
Sampling
Agentic behavior via recursive interactions
Elicitation
Server-initiated requests for user input - Server 主动请求用户输入
Roots
Filesystem boundaries for server access control
Summary
Protocol Architecture
MCP uses a two-layer architecture
| Layer | Desc |
|---|---|
| Data Layer | JSON-RPC 2.0 based communication with lifecycle management and primitives |
| Transport Layer | STDIO (local) and Streamable HTTP with SSE (remote) communication channels |
How MCP Servers Work
How MCP Servers Work
Request Flow
- A request is initiated by an end user or software acting on their behalf.
- The MCP Client sends the request to an MCP Host, which manages the AI Model runtime.
- The AI Model receives the user prompt and may request access to external tools or data via one or more tool calls.
- The MCP Host, not the model directly, communicates with the appropriate MCP Server(s) using the standardized protocol.
MCP Host Functionality
| Functionality | Desc |
|---|---|
| Tool Registry | Maintains a catalog of available tools and their capabilities. |
| Authentication | Verifies permissions for tool access. |
| Request Handler | Processes incoming tool requests from the model. |
| Response Formatter | Structures tool outputs in a format the model can understand. |
MCP Server Execution
- The MCP Host routes tool calls to one or more MCP Servers, each exposing specialized functions (e.g., search, calculations, database queries).
- The MCP Servers perform their respective operations and return results to the MCP Host in a consistent format.
- The MCP Host formats and relays these results to the AI Model.
Response Completion
- The AI Model incorporates the tool outputs into a final response.
- The MCP Host sends this response back to the MCP Client, which delivers it to the end user or calling software.
Real-World Use Cases for MCP
MCP enables a wide range of applications by extending AI capabilities:
| Application | Description |
|---|---|
| Enterprise Data Integration | Connect LLMs to databases, CRMs, or internal tools |
| Agentic AI Systems | Enable autonomous agents with tool access and decision-making workflows |
| Multi-modal Applications | Combine text, image, and audio tools within a single unified AI app |
| Real-time Data Integration | Bring live data into AI interactions for more accurate, current outputs |
MCP = Universal Standard for AI Interactions
- The MCP acts as a universal standard for AI interactions, much like how USB-C standardized physical connections for devices.
- In the world of AI, MCP provides a consistent interface
- allowing models (clients) to integrate seamlessly with external tools and data providers (servers).
- This eliminates the need for diverse, custom protocols for each API or data source.
- Under MCP, an MCP-compatible tool (referred to as an MCP server) follows a unified standard.
- These servers can list the tools or actions they offer and execute those actions when requested by an AI agent.
- AI agent platforms that support MCP are capable of discovering available tools from the servers and invoking them through this standard protocol.
Facilitates access to knowledge
- Beyond offering tools, MCP also facilitates access to knowledge.
- It enables applications to provide context to large language models (LLMs) by linking them to various data sources.
- For instance, an MCP server might represent a company’s document repository, allowing agents to retrieve relevant information on demand.
- Another server could handle specific actions like sending emails or updating records.
- From the agent’s perspective, these are simply tools it can use—some tools return data (knowledge context), while others perform actions.
- MCP efficiently manages both.
- An agent connecting to an MCP server automatically learns the server’s available capabilities and accessible data through a standard format.
- This standardization enables dynamic tool availability.
- For example, adding a new MCP server to an agent’s system makes its functions immediately usable
- without requiring further customization of the agent’s instructions.
- This streamlined integration aligns with the flow depicted in the following diagram
- where servers provide both tools and knowledge, ensuring seamless collaboration across systems.
Example: Scalable Agent Solution
- The Universal Connector enables MCP servers to communicate and share capabilities with each other
- allowing ServerA to delegate tasks to ServerB or access its tools and knowledge.
- This federates tools and data across servers, supporting scalable and modular agent architectures.
- Because MCP standardizes tool exposure, agents can dynamically discover and route requests between servers without hardcoded integrations.
- Tool and knowledge federation: Tools and data can be accessed across servers, enabling more scalable and modular agentic architectures.
Advanced MCP Scenarios with Client-Side LLM Integration
- Beyond the basic MCP architecture
- there are advanced scenarios where both client and server contain LLMs, enabling more sophisticated interactions.
- In the following diagram, Client App could be an IDE with a number of MCP tools available for user by the LLM:
Practical Benefits of MCP
Here are the practical benefits of using MCP
| Benefit | Desc |
|---|---|
| Freshness | Models can access up-to-date information beyond their training data |
| Capability Extension | Models can leverage specialized tools for tasks they weren’t trained for |
| Reduced Hallucinations | External data sources provide factual grounding |
| Privacy | Sensitive data can stay within secure environments instead of being embedded in prompts |
Key Takeaways
The following are key takeaways for using MCP:
- MCP standardizes how AI models interact with tools and data
- Promotes extensibility, consistency, and interoperability
- MCP helps reduce development time, improve reliability, and extend model capabilities
- The client-server architecture enables flexible, extensible AI applications
Exercise
Think about an AI application you’re interested in building.
- Which external tools or data could enhance its capabilities?
- How might MCP make integration simpler and more reliable?
All articles on this blog are licensed under CC BY-NC-SA 4.0 unless otherwise stated.
Related Articles
2024-07-06
LLM Core - Word2Vec
Word2Vec概述 在 NLP 中，在文本预处理后，进行特征提取，涉及到将词语转化成数值的形式，方便计算机理解 Word2Vec 的目的 - 将词语转换成向量形式，使得计算机能够理解 通过学习大量文本数据，捕捉到词语之间的上下文关系，进而生成词的高维表示 - 即词向量 架构 Word2Vec 有 2 种主要的模型 - Skip-Gram + CBOW Model Desc CBOW 根据周围的上下文词汇来预测目标词 Skip-Gram 根据目标词预测其周围的上下文词汇 优劣 Key Value 优点 揭示词与词之间的相似性 - 通过计算向量之间的距离来找到语义相近的词 缺点 无法处理多义词，每个词被赋予一个向量，不考虑上下文中的多种含义 模型架构 连续词袋 - Continuous Bag of Words, CBOW跳字模型 - Skip-Gram 连续词袋 CBOW 模型是一种通过上下文预测目标词的神经网络架构 上下文由目标词周围的一个词或多个词组成，这个数目由窗口大小决定 窗口是指上下文词语的范围 - 如果窗口为 10，那么模型将使用目标词前后各...
2024-08-13
RAG - Hybrid retrieval + Rerank
向量检索 当前主流的 RAG 检索方式主要采用向量检索，通过语义相似度来匹配 Chunk 向量检索并非万能，在某些场景下无法替代传统关键词检索的优势 当需要精准搜索的时候，向量检索的准确性就往往不如关键词检索 当用户输入的问题非常简短，语义匹配的效果可能不尽理想 关键词检索的适用场景 精确匹配 少量字符的匹配 - 不适合用向量检索 低频词汇的匹配 混合检索 结合关键词检索和语义匹配的优势 在 RAG 检索场景中，首要目标是确保最相关的结果能够出现在候选列表中 向量检索和关键词检索各具优势，混合检索通过结合多种检索技术，弥补各自不足，提供一种更加全面的搜索方案 重排序技术在检索系统中扮演着至关重要的角色 即使检索算法已经能够捕捉到所有相关的结果，重排序过程依然不可或缺 确保最符合用户意图和查询语义的结果优先展示，提升用户的搜索体验和结果的准确性 在多个数据集和多个检索任务中，混合检索和重排序的组合均取得了最佳表现 融合检索 / 多路召回 https://python.langchain.com/v0.2/api_reference/community/r...
2024-08-16
RAG - Evolution
演进 Naive RAG -> Advanced RAG -> Modular RAG 三个范式之间具有继承与发展的关系 Advanced RAG 是 Modular RAG 的一种特例形式 Naive RAG 是 Advanced RAG 的基础特例 RAG 技术不断演进，以适应更复杂的任务和场景需求 Naive RAG Naive RAG 是最基础的形式，依赖核心的索引和检索策略来增强生成模型的输出 Naive RAG 适用于一些基础任务和产品 MVP 阶段 Advanced RAG 通过增加检索前、检索中以及检索后的优化策略，提高检索的准确性和生成的关联性 - 适用于复杂任务 Advanced RAG 通过优化检索前、检索中、检索后的各个环节 在索引质量、检索效果以及生成内容的上下文相关性方面都取得显著提升 检索前 通过索引、分块、查询优化和内容向量化等技术手段，提高检索内容的精确性和生成内容的相关性 滑动窗口 overlap 经典的 Chunking 技术，通过在相邻的 Chunk 之间创建重叠区域，确保关键信息不会因简单的 Chunking 而丢失 在 ...
2024-07-04
LLM Core - RNN
背景 RNN 主要用来处理序列数据，目前大部分 LLM 都是基于 Transformer 通过学习 RNN，有助于理解 Transformer 有助于理解神经网络如何处理序列中的依赖关系、记忆过去的信息，并在此基础上生成预测 有助于理解关键问题 - 梯度消失 / 梯度爆炸 RNN Recurrent neural network - 循环神经网络 RNN 是一类用于处理序列数据的神经网络，RNN 能够处理序列长度变化的数据 - 文本 / 语音 RNN 的特点是在模型中引入了循环，使得网络能够保持某种状态，表现出更好的性能 左边 $x$ 为输入层，$o$ 为输出层，中间的 $s$ 为隐藏层，在 $s$ 层进行一个循环 $W$ 右边（展开循环） 与时间 $t$ 相关的状态变化 神经网络在处理数据时，能看到前后时刻的状态，即上下文 RNN 因为隐藏层有时序状态，那么在推理的时候，可以借助上下文，从而理解语义更加准确 优劣优势 RNN 具有记忆能力，通过隐藏层的循环结构来捕捉序列的长期依赖关系 特别适用于文本生成、语音识别等领域 局限 存在梯度消失和梯度爆炸的...
2024-08-14
RAG - LLM + Prompt Engineering
RAG 生成流程 经过 RAG 索引流程（外部知识的解析和向量化）和 RAG 检索流程（语义相似性的匹配及混合检索），进入到 RAG 生成流程 在 RAG 生成流程中，需要组合指令，即携带查询问题及检索到的相关信息输入的 LLM，由 LLM 理解并生成最终的回复 RAG 的本质是通过 LLM 提供外部知识来增强其理解和回答领域问题的能力 LLM 在 RAG 系统中起到了大脑的作用 在面对复杂且多样化的 RAG 任务时，LLM 的性能直接决定了系统的整体效果 提示词工程是生成流程中的另一个关键环节 通过有效的指令的设计和组合，可以帮助 LLM 更好地理解输入内容，从而生成更加精确和相关的回答 精心设计的问题提示词，往往能提升生成效果 LLM发展 RAG 目前更关注通用大模型 原理 Google 于 2017 年发布论文 Attention Is All You Need，引入了 Transformer 模型 Transformer 模型是深度学习领域的一个突破性架构，LLM 的成功得益于对 Transformer 模型的应用 与传统的 RNN（循环神经网络） 相比，Transform...
2024-07-01
LLM - API
背景 LLM 是没有 Web API 的，需要进行一次封装 将 LLM 的核心接口封装成 Web API 来为用户提供服务 - 必经之路 接口封装FastAPI接口封装 Uvicorn + FastAPI Uvicorn 类似于 Tomcat，但比 Tomcat 轻量很多，作为 Web 服务器 允许异步处理 HTTP 请求，非常适合处理并发请求 基于 uvloop 和 httptools，具备非常高的性能 FastAPI 类似于 SpringBoot，同样比 SpringBoot 轻量很多，作为 API 框架 结合 Uvicorn 和 FastAPI 可以构建一个高性能的、易于扩展的异步 Web 应用程序 Uvicorn 作为服务器运行 FastAPI 应用，可以提供优异的并发处理能力 安装依赖12$ pip install fastapi$ pip install uvicorn 代码分层1234567891011121314import uvicornfrom fastapi import FastAPI# 创建API应用app = FastAPI()@app.get(&quo...
