ARTICLE

Model Context Protocol (MCP)

Updated 2 May 2025
agentsmcpprotocoltoolsanthropicinfrastructure

Model Context Protocol (MCP)

MCP is the most important piece of infrastructure you’ve probably only just heard of. Created by Anthropic, it’s an open standard that solves a problem every AI developer hits: how do you connect AI to tools and data without building a custom integration for every combination?

The answer is: one protocol. Like USB replaced a dozen proprietary cables. Like HTTP gave us one way to fetch web pages regardless of the server. MCP gives AI models one standard way to access any tool, any data source, any service.

This matters because agents are only as useful as the tools they can reach.

The Problem MCP Solves

Before MCP, connecting AI to tools looked like this:

flowchart LR
    Claude --> CustomA[Custom Integration A]
    Claude --> CustomB[Custom Integration B]
    GPT --> CustomC[Custom Integration C]
    GPT --> CustomD[Custom Integration D]
    Gemini --> CustomE[Custom Integration E]
    Gemini --> CustomF[Custom Integration F]
    CustomA --> DB[(Database)]
    CustomB --> API[API]
    CustomC --> DB
    CustomD --> API
    CustomE --> DB
    CustomF --> API

Every model needed its own integration for every tool. N models × M tools = N×M custom integrations. Doesn’t scale.

With MCP:

flowchart LR
    Claude --> MCP[MCP Protocol]
    GPT --> MCP
    Gemini --> MCP
    MCP --> ServerA[MCP Server: Database]
    MCP --> ServerB[MCP Server: API]
    MCP --> ServerC[MCP Server: Files]
    MCP --> ServerD[MCP Server: Search]

N models + M servers = N+M implementations. Every model speaks MCP. Every tool exposes an MCP server. Universal compatibility.

How It Works

MCP uses a client-server architecture:

MCP Hosts

The applications that want to use AI with tools — Claude Desktop, Cursor, your custom app. They embed an MCP client.

MCP Clients

Protocol-level connectors that maintain connections to MCP servers. Handle the communication plumbing.

MCP Servers

Lightweight programs that expose tools and data. Each server provides:

  • Tools — Functions the AI can call (search, query, write, execute)
  • Resources — Data the AI can read (files, database records, API responses)
  • Prompts — Reusable prompt templates for common tasks

The Conversation

1. Client connects to Server
2. Server declares: "I offer these tools: [search, create_issue, read_file]"
3. AI decides it needs to search → "Call search with query='EU AI Act timeline'"
4. Client routes the call to the Server
5. Server executes the search, returns results
6. AI continues reasoning with the new information

The AI never talks to the tool directly. MCP handles the routing, authentication, and data formatting.

What You Can Build

MCP servers exist (or can be built) for essentially anything:

CategoryExamplesWhy It Matters
DatabasesPostgreSQL, SQLite, PineconeAI can query your data directly
DevelopmentGitHub, GitLab, JiraAI manages your repos and issues
KnowledgeGoogle Drive, Notion, ConfluenceAI reads your docs
CommunicationSlack, Email, DiscordAI can send and receive messages
SearchBrave, Google, internal searchAI gets real-time information
File systemsLocal files, S3, R2AI reads and writes files
Custom APIsYour internal servicesAI integrates with your stack

The power is composability. An agent with MCP access to your GitHub, your database, and your documentation can autonomously investigate a bug, find the relevant code, check the docs, and propose a fix.

Why This Changes Everything

Before MCP

Every AI tool integration was bespoke. Cursor had its own way of reading files. ChatGPT plugins had their own API spec. Every framework (LangChain, CrewAI, AutoGen) had its own tool interface. If you built a tool for one, you rebuilt it for the others.

After MCP

Build one MCP server. Every MCP-compatible client can use it. The ecosystem compounds — every new server benefits every existing client, and vice versa.

This is why the agent ecosystem is accelerating. The integration tax has dropped dramatically.

The Network Effect

MCP gets more valuable as more tools and clients adopt it. We’re in the early stages of this network effect. Anthropic open-sourced it deliberately — it only works as a standard if it’s universal.

Getting Started

Using MCP servers (easiest):

  1. Install Claude Desktop
  2. Add MCP servers to your config (database, files, GitHub, etc.)
  3. Claude now has access to those tools. Ask it to query your database or manage your repos.

Building an MCP server:

  • SDKs available in TypeScript, Python, Java, Kotlin, C#
  • A basic server can be built in under 100 lines
  • Publish it for the community or keep it internal

In your own applications:

  • Use the MCP client SDK to connect your AI application to any MCP server
  • Works with any LLM that supports tool use — not just Claude

What I’m Still Learning

  • How to design MCP servers that are intuitive for AI to use (tool naming, parameter design)
  • The security model — who controls what an AI can access through MCP?
  • How MCP will interact with competing standards (OpenAI’s approach, Google’s, etc.)
  • The governance of the standard as adoption grows

Go Deeper

Sources

enes