MCP prompts are reusable message templates an MCP server exposes through prompts/list and prompts/get. Here is how they work, with the 2026-07-28 changes.
The Model Context Protocol has three server features: prompts, resources and tools. This post covers prompts, which the specification describes as user-controlled.
MCP prompts are reusable message templates that a Model Context Protocol server exposes to its clients. A client lists them with prompts/list, fills one in with prompts/get and a set of arguments, and receives messages ready to send to a language model. The specification expects the user to pick a prompt, usually as a slash command.
Servers that offer prompts declare the prompts capability during initialization. The 2025-06-18 specification adds a listChanged flag, which tells clients whether the server will announce changes to its prompt list. The flow has three steps:
prompts/list. The call supports pagination with a cursor. Each prompt has a name, an optional title and description, and an optional list of arguments.prompts/get with the prompt's name and the argument values. The server returns a description and a list of messages, each with a role of user or assistant.listChanged, it sends notifications/prompts/list_changed.A prompts/get request from the specification looks like this:
{
"jsonrpc": "2.0",
"id": 2,
"method": "prompts/get",
"params": {
"name": "code_review",
"arguments": {
"code": "def hello():\n print('world')"
}
}
}Message content can be text, image, audio or an embedded resource. The specification says servers should return JSON-RPC error -32602 for an invalid prompt name or a missing required argument, and -32603 for internal errors. It also says implementations must validate prompt inputs and outputs to prevent injection attacks.
LeanIX Engineering's walkthrough from 26 November 2025 registers prompts on a FastMCP server with a decorator. The shortest form returns a string:
@mcp.prompt(
name="suggestActivity",
description="Returns a prompt to suggest activity",
)
def suggestActivity(type: str) -> str:
return f"Can you please suggest a '{type}' activity?"The walkthrough's version also sets tags and metadata, which are trimmed here. Its second example, generateCode, returns a PromptMessage instead of a string. FastMCP builds the argument list from the function signature, and the generated schema marks both language and task_description as required.
On the client side, the FastMCP prompts documentation shows client.get_prompt("user_greeting", {...}), which returns a GetPromptResult with the rendered messages. When a server exposes several versions of a prompt, the client takes the highest version by default and accepts a version argument to pick another.
The 2026-07-28 revision is the latest version of the specification. These changes touch prompts:
prompts/list results must carry ttlMs and cacheScope. The prompts concepts page shows "ttlMs": 600000 and "cacheScope": "public" in its example. The changelog describes ttlMs as a freshness hint in milliseconds, and cacheScope as public or private.resultType, which is complete for ordinary results.icons array, and prompt messages can carry a resource_link content type.prompts/get may return an InputRequiredResult when it needs more input before it can resolve the prompt. The client then retries the request with inputResponses.initialize handshake is removed. prompts/list results must not vary per connection, though they may vary by the authorization sent on the request.subscriptions/listen stream that the client opts into.Prompts are not on the 2026-07-28 deprecated features registry. That registry lists Roots, Sampling, Logging, Dynamic Client Registration, the HTTP+SSE transport and two includeContext values. For Roots, Sampling and Logging, the earliest removal is the first revision released on or after 2027-07-28.
Use a prompt when a user should choose a workflow on purpose and the wording should stay the same each time, such as a code review template or a standard report request. The specification treats prompts as user-selected, so a prompt does not suit a step the model should decide to take on its own.
Client support is uneven. The LeanIX walkthrough says it plainly: "Not all MCP clients support every feature, and some may have limitations when listing certain tools or handling specific functionalities." Test a prompt in the client you target before you build a flow around it.
prompts/list and rendered with prompts/get.name, title, description and arguments for each prompt.ttlMs, cacheScope, resultType and icons, and lets prompts/get ask for more input.No. The specification describes prompts as user-controlled, meaning the user decides when one is used. Implementers may expose them as slash commands or through any other interface they choose.
Each argument has a name, an optional description and a required flag. Servers should validate arguments before they process them, and a missing required argument should return JSON-RPC error -32602.
No. The 2026-07-28 deprecation registry does not list prompts. It lists Roots, Sampling, Logging, Dynamic Client Registration, HTTP+SSE and two includeContext values.
To compare MCP servers before you wire one into a client, browse the MCP servers on aiagentslisting.com.
One email a week. New agents, MCP servers and skills, and what is actually getting traction.
A Copenhagen agency founder on logging MCP agent runs, splitting agent time from human review, and turning a run into an invoice line without watching anyone's screen.
5 min read
MCP elicitation lets a server pause a tool call to ask the user for missing input, in form mode or URL mode, before it continues. Here's how it works.
1 min read
MCP sampling lets a server ask the client's model for a completion. The 2026-07-28 MCP spec deprecates it, so here is how it worked and how to migrate.
4 min read