Prompts

The third primitive

Tools do things, resources hold data, and prompts are the third piece: reusable, parameterized templates that kick off a particular kind of interaction. Think of them as a server-hosted version of the few-shot and templated prompts from the LangChain course, just discoverable and fillable by a client instead of hardcoded into your own script.

@mcp.prompt()
def code_review(language: str, code: str) -> str:
    """Ask for a structured code review of a snippet."""
    return (
        f"Please review this {language} code for bugs, style issues, "
        f"and possible improvements:\n\n{code}"
    )

Discovering and filling in a prompt

Same list/get shape again:

• prompts/list (await mcp.list_prompts()): what prompt templates exist?
• prompts/get (await mcp.get_prompt(name, arguments)): fill in this template's blanks and hand back the resulting message(s).

prompts = await mcp.list_prompts()
# [Prompt(name='code_review', description='Ask for a structured code review of a snippet.', ...)]

result = await mcp.get_prompt("code_review", {"language": "python", "code": "print(1)"})
# result.messages -> [PromptMessage(role='user', content=TextContent(text='Please review...'))]

Notice the return shape: a list of messages, not a single string. A prompt template can hand back a whole conversation starter, a system message plus an example user turn, not just one line of text.

Why a server would bother with this

A server author who knows their domain well, say, a code-review tool provider, can ship the right way to ask for a review alongside the tool that performs one. A host application can surface these as "slash commands" or a menu a user picks from, rather than the user needing to know how to phrase the request themselves. A well-designed prompt gets better results than an ad hoc one, and here that expertise ships with the server instead of living in your own codebase.