What is MCP (Model Context Protocol), and why does it matter?

The problem it solves: before MCP, every AI app wrote custom integrations for every tool. With 10 apps and 50 tools, that's up to 500 integrations.

Core concepts

• Host / client: the AI application that connects to servers.
• Server: exposes capabilities. It offers three main primitives:
  • Tools: actions the model can call (model-controlled).
  • Resources: data the app can read into context, like files or DB rows (application-controlled).
  • Prompts: reusable prompt templates (user-controlled).
• Transports: stdio for local servers, and HTTP-based transports (streamable HTTP) for remote servers.
• Messages use JSON-RPC 2.0.

MCP is not a replacement for function calling. Function calling is how the model requests an action. MCP is how the application discovers and connects to the tools that get exposed to the model.

Concerns to raise (shows maturity)

• Security: a malicious or compromised MCP server can inject instructions through tool descriptions or results ("tool poisoning"). Vet servers, pin versions, scope permissions.
• Tool overload: connecting many servers can expose hundreds of tools, which hurts tool selection (see Q19).
• Auth: remote servers need proper OAuth and per-user credentials, not shared keys.

Follow-ups to expect

• MCP vs A2A (agent-to-agent protocols)? MCP connects agents to tools and data. A2A-style protocols connect agents to other agents.

Related

• Previous: Q6. Explain short-term vs long-term memory in agents.
• Next: Q8. How do you write a good tool definition for an agent?
• Model Context Protocol