Creativity · MCP — pattern
MCP Prompts Capability: Deep Dive
Alongside tools and resources, the Model Context Protocol defines a 'prompts' capability: a server can publish a list of named prompt templates with typed arguments, and a client can render these as slash commands, quick actions, or library entries. This page unpacks how prompts are declared, listed, and executed — plus where the capability shines versus where you should use a tool instead.
MCP facts
- Kind
- pattern
- Ecosystem
- anthropic-mcp
- Transports
- stdio, sse, http
Capabilities
- prompts/list — client enumerates all prompt templates exposed by the server
- prompts/get — client fetches a rendered prompt with provided arguments
- Typed arguments: each prompt declares name, description, and required params
Frequently asked questions
When should I use a prompt vs a tool?
Prompts render text the user wants to send to the model — 'summarize this PR', 'review these diffs'. Tools perform side effects — 'create issue', 'write file'. If the server action has no side effect beyond producing a prompt, use a prompt.
Do all clients support prompts?
Claude Desktop, Cursor, Zed, Continue, and Windsurf all honor prompts. Simpler clients may ignore the capability; servers should treat prompts as optional sugar rather than a hard dependency.
Can prompts reference resources?
Yes — a rendered prompt can embed @resource references that the client resolves before sending to the model. This keeps long documents out of the prompt text itself.
Are prompt arguments typed?
Yes — each argument declares name, description, and required flag. Richer schema support is an active area of the MCP spec; treat complex validation as best-effort today.
Sources
- MCP Prompts spec — accessed 2026-04-20
- MCP docs — accessed 2026-04-20