MCP: One Protocol for All AI Connections

Model Context Protocol (MCP) is a standard client–server architecture that lets AI applications connect to external systems to perform actions, access data, and reuse prompt templates.

Instead of building one-off integrations for every tool or data source, MCP creates a consistent way for AI hosts and external services to communicate.

MCP standardizes how AI apps access actions, data, and prompt templates through a simple client–server architecture.

Servers can expose three primitives. Each has a distinct purpose and is used differently by the AI.

Protocol & Transport

• Protocol: JSON-RPC 2.0 for initialization, discovery, requests, responses, and notifications
• Transport: Message delivery only, using STDIO for local scenarios and HTTP + SSE for remote ones

Lifecycle Flow

1. Client connects to server
2. Initialization handshake
3. Server advertises capabilities
4. Client uses tools / resources / prompts
5. Results (and optional notifications) returned
6. Connection closes

• The AI Host controls which servers are connected and when tools are executed.
• The MCP Server responds only when called; it does not control the model or initiate actions.
• Tool execution can require user approval depending on host policy.

Consolidated Minimal View

AI Host
  └─ MCP Client
       └─ MCP Server
            ├─ Tools      (Actions)
            ├─ Resources  (Context)
            └─ Prompts    (Instructions)