The Complete Guide to Model Context Protocol (MCP): Architecture & Real-World Use Cases

As AI models like Claude, GPT-4, and open-source LLMs evolved, developers faced a repeating problem: every AI assistant required custom integrations, API wrappers, and prompt hacks to read files, run database queries, or check project specs.

Model Context Protocol (MCP) solves this by providing an open standard for bidirectional communication between an AI application (the Host) and an external context provider (the Server).

MCP operates on a client-server model over JSON-RPC 2.0. The architecture consists of three key entities:

• Host Application: The AI program the developer interacts with (e.g. Claude Code CLI, Cursor IDE, Windsurf).
• MCP Client: The protocol implementation inside the Host that initiates requests and handles protocol negotiation.
• MCP Server: The service exposing context, executable tools, or prompt templates to the AI client.

Here is how a standard JSON-RPC tool invocation request looks over MCP:

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "get_feature_spec",
    "arguments": {
      "feature_id": "F-102"
    }
  }
}

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "content": [
      {
        "type": "text",
        "text": "Feature F-102: User Authentication. Rules: 1. Require 2FA for admin roles. 2. Tokens expire in 24 hours."
      }
    ]
  }
}

MCP divides functionality into three fundamental primitives:

1. Tools (Executable Actions)

Tools allow AI agents to perform side effects or query dynamic data (e.g. running tests, creating database migrations, reading a feature specification).

2. Resources (Passive Context Data)

Resources act like file handles or URI-addressable state (e.g. file:///project/schema.sql or contextsbase://features/F-102). AI models can read resources to load context into their prompt window.

3. Prompts (Reusable Templates)

Prompts are pre-configured workflow templates exposed by the server to guide the AI assistant on specific engineering tasks.

MCP supports two main transport layers for message delivery:

• Codebase Memory & Specs: Serving architectural rules, entities, and specs to AI coding agents so they write compliant code without hallucinating.
• Live Database Querying: Allowing agents to inspect SQL schemas and safely execute read-only queries during debugging.
• API Inspection & Testing: Interacting with local environment APIs, Postman collections, and automated test runners.
• Production Monitoring & Logs: Feeding Sentry errors or Datadog log snippets directly into the agent's context during root-cause analysis.

ContextsBase operates as a native HTTP SSE MCP server. Instead of letting coding agents start every task from zero context or rely on static markdown files, ContextsBase serves living features, entity data models, business rules, and test suites over MCP directly to Claude Code, Cursor, or Copilot.