MCP prompt
A reusable message template an MCP server offers, with named arguments, for a person or application to pick, rather than for the model to call.
The specification calls prompts "user-controlled": they are meant for a person to choose explicitly, often as a slash command. They let a server's author ship the instructions that work best with their server, such as a review checklist or a triage format, so every user does not have to write their own.
A prompt has a name and, optionally, a title, a description and a list of arguments. Each argument has a name and may have a description and a required flag. The client lists prompts with prompts/list and fetches one with prompts/get, passing the arguments. The server returns messages: each has a role (user or assistant) and content, which can be text, an image, audio or an embedded resource. A server that declares listChanged sends notifications/prompts/list_changed when its prompts change.
Example
{ "jsonrpc": "2.0", "id": 2, "method": "prompts/get",
"params": { "name": "code_review",
"arguments": { "code": "def hello():\n print('world')" } } }The server answers with one user message, "Please review this Python code: ...", ready to send to a model. An unknown prompt name or a missing required argument is a -32602 error.
Common confusion
A prompt is not a tool. The model never calls a prompt; a person or the application picks it, and its text becomes part of what the model reads. That is also the risk: prompt text comes from the server and goes straight to your model. The spec says implementations "MUST carefully validate all prompt inputs and outputs to prevent injection attacks", and a server that quietly changes a prompt can steer your model the way tool poisoningTool poisoning: An attack that hides instructions for the model in a tool's description, schema or results, so the agent follows them while the user sees only a harmless tool. does.
In L2M
Agent, Supervisor, LLM Call and chat model nodes have a Server prompt picker. The workflow builder picks a registered server's prompt, fills its arguments (templates allowed) and chooses whether it replaces the node's system or user message. L2M fetches the prompt each time the node runs, so the server's current version applies, and records what it used under _serverPrompt in the node's output (using a server prompt). The messages become one text: assistant turns are labelled, and images or binary parts are replaced by a short placeholder. Optional tool policy rules of kind prompt can deny prompts by name (rule kinds). The MCP gateway serves tools only, not prompts.
Further reading
- MCP specification 2025-11-25: prompts (modelcontextprotocol.io)