What are MCP prompts?

MCP prompts are ready-made request templates from an MCP server. You pick one, fill in the blanks and run it.

MCP prompts are reusable instruction templates an MCP server offers, like a ready-made request that you fill in and run. You choose them yourself, often as slash commands. The AI does not pick them.

Also called MCP prompt, prompts/list, prompts/get, MCP prompt templates, Model Context Protocol prompts.

A prompt can take arguments, so one template covers many requests. The docs give examples such as plan a vacation, summarize my meetings or draft an email. When you run one, the server returns structured messages, which the app adds to your conversation.

The spec says prompts are user-controlled. You explicitly select them, and typically through a command in the interface, such as a slash command. The server writes the content of the prompt, but you decide when it runs.

Prompts are one of three things a server can offer. The others are tools (actions the AI chooses) and resources (context the app chooses).

MCP prompts are ready-made request templates an MCP server offers, which you pick and run with your own details filled in. The one thing to know first: the server writes the prompt text, and the spec says inputs and outputs must be validated against injection.

In Claude Code, type / to see them. Each MCP prompt is listed as /servername:promptname (MCP), and typing /mcp__servername__promptname also runs it. Arguments go after the command, separated by spaces. The result is injected into the conversation.

Cursor's docs also list prompts as supported. Check your own app's docs for the rest. The MCP docs say tools can write to databases or modify files, while prompts are instruction templates. See MCP tools.

Good for

Claude Code users who see server commands in the slash menu
Teams that want a server to ship its own best-practice requests
Builders choosing between a prompt, a tool and a resource

Comparison

A prompt is a shortcut you start. A tool is something the AI decides to run.

Option	What it is	Who decides	Can change things
PromptsThis page	Instruction templates	You	Not applicable
Tools	Actions	The AI model	Yes
Resources	Context data	The app	No

Questions

What are MCP prompts?

Ready-made instruction templates that an MCP server offers. You pick one, give it any details it asks for and run it. The server writes the text.

How do I use MCP prompts in Claude?

In Claude Code, type / and pick the server's prompt from the list, shown as /servername:promptname (MCP). Add any arguments after it. Cursor's docs list prompts as supported too. For other apps, check your app's docs.

What is the difference between MCP prompts and tools?

You choose a prompt, and it supplies an instruction template. The AI model chooses a tool, and the MCP docs say tools can write to databases, send messages or modify files.

Who controls an MCP prompt, me or the AI?

You. The spec calls prompts user-controlled, meant to be selected explicitly. The server writes the content, but you decide when it runs.

Can a prompt take inputs?

Yes. Prompts can have arguments, and the app can suggest values for them. In Claude Code you type the arguments after the command.

Are MCP prompts safe?

The text comes from the server, and the spec says inputs and outputs must be validated to prevent injection. Our advice: read what a prompt adds to your chat before you send it, and only use prompts from servers you trust.

Why don't I see my server's prompts?

Check that your app supports prompts and that the server offers any. In Claude Code, a server named anthropic-skills is hidden because the name is reserved. Rename the server to list its prompts.

Are MCP prompts the same as Claude skills?

No. A prompt comes from an MCP server and runs when you pick it. Agent skills are a separate feature. See agent skills.

Sources

Docs, issue trackers and discussions this page is built from.

Show all 4

Prompts · modelcontextprotocol.io
Understanding MCP servers · modelcontextprotocol.io
Connect Claude Code to tools via MCP · code.claude.com
Model Context Protocol (MCP) | Cursor Docs · cursor.com