MCP is an open protocol that revolutionizes how AI applications and agents interact with tools and data sources. Let’s dive into what makes it special.

MCP Architecture Diagram

MCP’s architecture consists of two main components:

  • Clients: These host AI applications
  • Servers: Provide federated access to systems and tools relevant to specific tasks

Key Components

MCP operates through three primary mechanisms:

  1. Tools (Model Controlled)

    • Controlled by the model
    • Model decides when to invoke functions to take actions

  2. Resources (Application Controlled)

    • Represents data exposed to applications
    • Includes files, databases, and other data sources

  3. Prompts (User Controlled)

    • User invokes prompt templates (e.g., slash commands)

What Makes MCP Special?

  • Dynamic Nature: Resources and prompts can be dynamic, allowing applications to subscribe to them. When changes occur, the application gets notified to take follow-up actions.

  • Context Management: MCP’s vision extends beyond just providing context to the LLM - it’s about what context is given to the application itself.

  • Agent Framework Integration: While agent frameworks design how the LLM performs in a loop and determine context usage, MCP protocol facilitates bringing that context to the LLM.

Real-World Applications

Several exciting implementations and tools have emerged:

  • Zapper MCP: Bringing new capabilities to the ecosystem
  • Remote MCP: Extending functionality across distributed systems
  • MCP Think Tool: Enhancing thinking capabilities
  • MCP Inspector: Testing MCP servers

MCPDoc Server

A notable implementation by LangChain that:

  • Provides MCP host applications (like Cursor, Windsurf, Claude Code/Desktop) with:
    1. User-defined list of llms.txt files
    2. Simple fetch_docs tool to read URLs within provided llms.txt files
  • Enables audit capabilities for tool calls and returned context
  • Can be used with tools like llmstxt_architect for creating llms.txt files

This protocol represents a significant step forward in AI application integration, offering a structured way to manage model-tool interactions while maintaining flexibility and control across different components of the system.

[update - 4/2/25] The easiest way to build an MCP Server using @GoogleDeepMind

  1. Use Gitingest to get all the code and docs from the FastMCP repo
  2. Download the code into a txt file
  3. Go to AI Studio, upload the file, define what kind of MCP Server you want to build
  4. Gemini 2.5 Pro builds it for you.

[other resources] Highly recommended by Neils Rogge “MCP crash course for Python developers” - https://x.com/NielsRogge/status/1913679695192670392?t=dC5024fM34pHDzuUdmfYSg&s=08 MCP 101 AI Engineering Academy