leejunkim

Model Context Protocol (MCP)

4 min read

Mcp

An open standard developed by Anthropic to enable large language models (LLMs) to interact seamlessly with external tools, systems, and data sources

  • Allows access to various external systems (computer files, an API, a browser, or anything else)
  • reading files, executing functions, querying databases, and interacting with APIs → makes them more “context aware”

Introduction

  • Acts as a “universal translator” or “USB-C port”
    • Addresses the fundamental challenge of AI models needing consistent and secure access to necessary information, regardless of location
  • The emergence of MCP reflects an important trend of standardization in the AI tool ecosystem
    • Just as how APIs standardized backend, MCP has the potential to revolutionize how AI applications interact with diverse data and tool environments
    • Ultimately, MCP is seen as a necessary evolutionary step for the AI ecosystem to mature and efficiently scale

Insights

  • A core idea of agentic AI, where LLMs don’t just simply generate text, but has a clear intention to do specific processes/tasks to achieve a goal
  • Solves the problem of LLMs of having knowledge limitations
    • MCP acts as a perfect way to get access to “updated” knowledge and act actively as an AI agent
  • By separating AI models (clients) from data/tool providers (servers), developers can independently develop, update, and scale these components
    • a fundamental architectural pattern for building robust enterprise-grade systems
      • For example, separate teams can independently develop MCP servers for financial data and CRM data without affecting the core AI application

MCP architecture

Fundamentally follows a client-server architecture

  • this directly supports modularity and scalability of AI systems

MCP hosts

  • An AI application that receives user requests and seeks to access context through MCP - the main application
  • Examples
    • Claude for desktop, GPT Desktop, Cursor, different LLM applications that can connect to MCPs
  • tools that want to access data through MCP
  • LLM in here decides if MCP is necessary
    • If so, the host uses the MCP client as the orchestrator

MCP client

  • Resides within the MCP Host and handles communication between the host and the server
  • transforms user requests into a structured format that the MCP protocol can process
    • responsible for sending the request to the appropriate MCP server
  • session manager
    • handling interruptions, timeouts, reconnections, and session terminations
    • parses responses, handles errors, and ensures that responses are appropriate for the context
  • Protocol clients that maintain 1:1 connections with servers
  • client can connect to multiple servers
  • Examples
    • IBM® BeeAI, Microsoft Copilot Studio, and Claude.ai

MCP server

  • just think of them as “plugins”<<
  • Provides standardized access to specific data sources (e.g., GitHub repositories, Slack workspaces, AWS services, financial databases
  • Resources: File-like data that can be read by clients (like API responses or file contents)
  • Tools: Functions that can be called by the LLM (with user approval)
  • Prompts: Pre-written templates that help users accomplish specific tasks

Example Interaction Walkthrough

  1. The user interacts with the MCP host application
  2. The LLM within the host determines that external data or tools are needed
  3. The MCP host uses the MCP client to translate this need into a tool call of the MCP protocol
  4. The MCP client routes the request to the appropriate MCP server
  5. The MCP server executes the requested tool or retrieves the data
  6. The MCP server returns the formatted results to the MCP client
  7. The MCP client passes this result back to the host/LLM to be used in generating the final response

Advantages

  • Standardization and Interoperability
    • MCP provides a common language for AI to connect with data and tools, enabling seamless interaction between different systems (not tied to specific vendors/technologies)
  • Simplified Integration and Reduced Overhead
    • eliminates the need to develop custom connectors for every tool and data source
  • Enhanced Security and Governance
    • Existing security mechanisms (e.g., AWS IAM) can be utilized to apply consistent access control
    • Authentication
  • Scalability and Configurability
    • As MCP is based on a modular design, it supports the construction of scalable AI solutions that align with AWS architecture best practices
  • Context-Aware AI
    • AI models can utilize richer and more accurate contextual information

Resources

Setting MCP (claude) on my pc

4/24/25

Graph View

Backlinks