MCPBeginner

What MCP solves

“What problem does MCP solve compared with implementing every integration directly?”

What this tests

  • Understanding the M×N integration problem and the client-server split
  • Knowledge of MCP primitives: tools, resources, prompts, discovery
  • Realistic view of costs: process boundary, auth, trust in third-party servers
  • Whether they can say when direct integration is still better

Answers by level

Read the beginner answer first and notice what is missing.

Without a protocol, every agent host has to implement every integration: M hosts × N services, each with its own tool definitions, auth handling, and result formatting. MCP splits this into a client (the host application) and servers that expose capabilities in a standard shape: tools (callable functions with JSON schemas), resources (readable data), and prompts (templates), with runtime discovery so the client learns what a server offers instead of hardcoding it. See MCP Overview and MCP Primitives: Tools, Resources, Prompts.

The practical win is reuse and decoupling: a GitHub server written once works in any MCP-capable host; the integration lives in its own process with its own credentials, and can be updated independently. It also standardises the lifecycle: initialise, capability negotiation, tool listing, invocation, and errors. See MCP Lifecycle: Init, Discovery, Invocation.

It does not make the model better at using tools. The model still sees tool names, descriptions, and schemas, and all the tool-design rules apply. It adds a process boundary and transport, which is latency and one more thing to operate.

Green flags · Red flags

Green flags
  • M×N framing, client-server split, discovery
  • Names tools, resources, prompts
  • States that MCP does not improve tool selection by itself
  • Mentions trust and auth for third-party servers
  • Says when direct integration is the better choice
Red flags
  • Describes MCP as making agents smarter
  • No awareness of the security surface of external servers
  • Cannot compare with plain function calling
  • Suggests connecting every available server

Follow-up questions

F1
A server exposes 60 tools. What do you do?
F2
Where do credentials live?

Practical scenario

Your company has an internal ticketing API. Three teams want to use it from three different agent hosts. One team proposes writing an MCP server; another argues for a shared TypeScript library with function-calling definitions. Lay out the tradeoffs and recommend one, including how you would secure it.

Related concepts · Learn this topic

MCP OverviewMCP Primitives: Tools, Resources, PromptsMCP vs Direct Integration vs Function CallingMCP Lifecycle: Init, Discovery, Invocation
← Why more memory is not betterMCP vs function calling →