Concept guide · Agents, tools, and MCP
What is MCP and how is it different from tool calling?
A slower explanation for when the short interview answer is not enough. Start with the intuition, then connect it to the technical idea and the way an interviewer expects you to reason about it.
ML area
Generative AI → MCP and tool calling
Interview format
Core · Scenario
Question
GENAI-080
Start with the idea
In simple terms
Tool calling is the model asking an application to run a named action with arguments. MCP is a standard way for compatible clients to discover and use tools, resources, and prompts exposed by a server.
Make it concrete
A model can request `search_orders` through a tool schema. An MCP server can expose that capability, along with related resources, to more than one AI application using a common protocol.
Build the answer
How to approach it
- 1Separate the model’s requested action from the protocol and the actual execution.
- 2Define trust, authentication, permissions, and tenant boundaries for each server.
- 3Test malformed arguments, malicious results, timeouts, and revoked access.
Go one level deeper
The technical explanation
The protocol does not make a tool trustworthy. The host or client still decides which servers are connected, the server still needs authentication and authorisation, and the application must validate arguments and results. Explain the roles, host, client, and server, then discuss lifecycle, versioning, timeouts, result size, and prompt or tool-description injection.
Say this in the interview
Tool calling is a model interaction in which the model proposes a named action with arguments. MCP is a protocol for exposing tools, resources, and prompts through a standard server-client arrangement so compatible applications can connect to capabilities consistently. Both still require authorisation and validation.
Keep preparing