How does the MCP Prompts primitive differ from the Tools primitive in terms of who controls invocation?

MCP servers expose three kinds of capability: tools, resources, and prompts. The exam question is about who pulls the trigger on each. **Tools** are model-initiated. The LLM sees a list of tool descriptions in its context, and when the conversation calls for one, the model decides on its own to call it with arguments. Nobody clicks a button; the model just acts. **Prompts** are user-initiated. The server defines a reusable, parameterized message template, and the host application surfaces it in its UI, very often as a slash-command you type. The user picks the prompt, fills in any parameters, and that populated message gets sent into the conversation. So the difference is the initiator. Tools answer 'the model wants to do something'. Prompts answer 'the user wants to kick off a known workflow'. Both can carry parameters, so 'prompts are static text' is wrong. The real split is control: model-controlled versus user-controlled.

The difference is who initiates invocation. That single axis is what the question tests. **Tools** are model-controlled. The host registers each tool's schema into the model's context. The LLM reads those descriptions and autonomously decides when to call one, supplying arguments. The user never explicitly selects a tool. The model does, based on conversation context. **Prompts** are user-controlled. An MCP server defines a named, parameterized message template. The host surfaces it in a UI affordance, most commonly a slash-command or a template picker. The user explicitly chooses the prompt and fills in its parameters. The populated messages then enter the conversation. This maps to a clean three-way control model in MCP. Tools are model-controlled. Prompts are user-controlled. Resources are application-controlled, meaning the host decides what context to attach. Separating these is deliberate. It keeps autonomous actions and human-triggered workflows on different rails. The distractors fail predictably. Both being model-invoked erases the whole point. Servers do not invoke prompts. And prompts do support parameters, so the static text claim is false.