What is an MCP prompt?

MCP prompt, explained#

Prompts are the third MCP primitive and the specification calls them user-controlled. A tool is chosen by the model and a resource is attached by the application, but a prompt is something the person deliberately invokes, the way you would pick a saved workflow.

A prompt has a name, an optional title and description, and a list of arguments, each with a name, a description and whether it is required. When the client calls prompts/get with values for those arguments, the server returns a list of messages, each with a role and content. The content can be text, or embed resources so the prompt arrives with its context.

Prompts are where a server author's knowledge goes. The author knows which tools to call first, what to check before changing anything, and where to stop and ask the person. Writing that into a prompt gives every user the same reliable workflow, instead of each one discovering it by trial and error.

Because prompts only produce messages, they run nothing on their own. The model still has to call tools, and the host still asks for consent for each call. Servers can change their prompt list during a session and notify clients, so the menu a person sees can depend on what they are allowed to do.

For example, a server offers weekly_review with an optional team argument. A lead types /weekly_review in their assistant and picks a team. The returned prompt tells the model to fetch what shipped, what slipped and what is blocked, and to draft a summary without posting it.