MCP Purpose & Scope

Why MCP exists

Before MCP, every AI application had to build custom integrations for every data source or tool it wanted to use — the N×M integration problem: N applications × M integrations, each one bespoke.

MCP solves this by defining a single open standard for connecting AI applications to external systems. Build an MCP server once, and any MCP-compatible host can use it.

flowchart LR
subgraph before["Before: N×M custom integrations"]
  A1[App 1] --> D1[(Data source 1)]
  A1 --> D2[(Data source 2)]
  A2[App 2] --> D1
  A2 --> D2
  A3[App 3] --> D1
  A3 --> D2
end
subgraph after["With MCP: N+M"]
  B1[App 1] --> P{{MCP}}
  B2[App 2] --> P
  B3[App 3] --> P
  P --> S1[(Server 1)]
  P --> S2[(Server 2)]
end
before ~~~ after

N×M custom integrations vs. one MCP standard

Think of MCP as the “USB-C of AI applications”: one standard connector between models and the outside world.

What MCP standardizes (in scope)

Context exposure

How context (data, tools, prompts) is exposed to AI applications.

Message format

JSON-RPC 2.0 messages exchanged between clients and servers.

Version & capability negotiation

How clients and servers agree on protocol versions and features — a connection handshake in legacy revisions; per-request metadata plus server discovery in the current stateless revision.

Primitives & transports

Server primitives (tools, resources, prompts), client primitives (e.g. elicitation), and standard transports: stdio and Streamable HTTP.

Tip

Primitives evolve with the spec: sampling and roots exist as client primitives in earlier revisions but are deprecated in the latest one. Always check the spec revision your exam targets.

What MCP does NOT standardize (out of scope)

• Which AI model is used, or how the model reasons
• How the host application renders UI or manages conversations
• The internal implementation of servers (any language, any backend)
• Agent orchestration logic (MCP provides building blocks, not an agent framework)

Key exam points

• MCP is open source and model-agnostic — it works with any LLM, not just Claude
• MCP was introduced by Anthropic in November 2024; in December 2025 Anthropic donated it to the Agentic AI Foundation (AAIF), part of the Linux Foundation
• MCP standardizes the interface, not the implementation
• MCP reduces integration cost from N×M to N+M

Check your knowledge

📝 Check your knowledge

1. What integration problem does MCP solve?

ASlow LLM inference timesBThe N×M problem of building a custom integration for every app–data source pairCThe lack of a standard prompt-engineering formatDIncompatibility between different LLM vendors' model weights

2. Which of the following is IN scope for the MCP specification?

AHow the host renders its user interfaceBWhich AI model the host must useCThe JSON-RPC 2.0 message format between clients and serversDThe programming language servers must be written in

3. Is MCP tied to a specific AI model or vendor?

AYes, it only works with ClaudeBYes, it requires Anthropic-hosted modelsCNo, it is an open, model-agnostic standardDNo, but it only supports open-weight models

4. Who governs MCP today?

AAnthropic alone, as a proprietary protocolBThe Agentic AI Foundation (AAIF), part of the Linux Foundation, after Anthropic donated it in December 2025CThe W3CDOpenAI, via the AGENTS.md working group

5. Which pair correctly matches MCP primitives to the side that provides them?

ATools and prompts are client primitives; elicitation is a server primitiveBTools, resources, and prompts are server primitives; elicitation is a client primitiveCResources are a transport; tools are a message formatDAll primitives are provided by the host application