Quartz 4

Model Context Protocol (MCP)

7 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
    • MCP lets an AI assistant talk to various software tools using a common language, instead of each tool requiring a different adapter or custom code
  • 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
  • A core idea of agentic AI
    • LLMs don’t just simply generate text, but has a clear intention to do specific processes/tasks to achieve a goal
  • 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

The problems it solved

  • Problem of fragmented integrations
    • Each integration needs a custom adapter (it’s brittle and doesn’t scale)
    • Ex) AI IDE might use one method to get code from Github, another to fetch data from a DB, etc
    • MCP offers one common protocol ⇒ Instead of writing separate code for each tool, a developer can implement the MCP specification and instantly make their application accessible to any AI that speaks MCP
  • 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

Basically (from what i understood) 📌

  • Before and after
    • Before every AI app had to build custom code to each tool (like Google drive)
    • But now we just get the server someone else made for us and just “plug it in” like a USB (users POV)
      • If you’re a a developer, you can can implement the MCP specification and instantly make their application accessible to any AI that speaks MCP
  • In order to build an MCP server
    • You have to expose the endpoints to the LLM with good documentation (input/output) for the LLM to understand
      • Make it easy for the AI to use
    • You have to understand the app pretty well & know how to use it programmatically
  • Use cases
    • Unity MCP
      • Rapid prototyping
      • Making LLM do the boilerplate of setting the environment (setting 100 trees in random locations)
      • Doing automated AI testing (edge cases, like what happens if I put the character there? if i do xyz?)
    • Figma MCP

MCP architecture

  • 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
    • 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 (the 1:1 connection) to an MCP server (or more)
      • If the AI wants to use a particular tool, it will connect through an MCP client to that tool’s MCP server
      • The client will then handle the communication (open a socket, send/receive msg) and present the server’s responses to the AI model
    • 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
    • Examples
      • IBM® BeeAI, Microsoft Copilot Studio, and Claude.ai
  • MCP server
    • Lightweight adapters that run alongside a specific application or service, that exposes that application’s functionality (its “services”) in a standardized way
      • just think of them as “plugins”<<
      • Ex) a Blender MCP server knows how to map “create a cube and apply a wood texture” into Blender’s Python API calls
    • The server primitives
      • 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
  • The MCP Protocol
    • The language and rules that the clients/servers use to communicate
    • Defines message formats, how a server advertises its available commands, how an AI asks a question or issues a command, how results are returned, etc
    • Transport-agnostic
      • The protocol can work over HTTP/Websockets for remote/standalone servers, or even standard IO streams for local integrations
    • The protocol ensures that whether an AI is talking to a design tool or a db, the handshake and query format are consistent
      • This consistency allows AI to use one MCP server to another (“grammar” remains the same)

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

Questions

  • How does MCP extend beyond the capabilities of a typical Retrieval-Augmented Generation (RAG) system?
    • While RAG provides passive context by retrieving text, MCP allows the model to actively fetch context or perform actions by invoking tools.
  • What is meant by ‘dynamic discovery’ in the context of MCP?
    • AI agents can automatically detect available MCP servers and their capabilities at runtime without needing hard-coded integrations.
  • What is a key security principle for MCP implementors regarding tool usage?
    • Hosts must obtain explicit user consent before invoking any tool, and users must understand what the tool does.
  • Hosts must obtain explicit user consent before invoking any tool, and users must understand what the tool does.
  • When might MCP be considered overkill for an application?
    • If an AI model only needs to access one or two straightforward APIs, direct API calls might be a more efficient solution.

Graph View

Backlinks