What it does
The client sends a prompts/get request with a required name field identifying the prompt and an optional arguments map of string key-value pairs. Before sending the request, the client must confirm the server declared a prompts capability during initialize negotiation — if the server omitted that capability block, prompts/get will return a JSON-RPC -32601 MethodNotFound error. The server looks up the named prompt template, interpolates the provided arguments into the template slots, and synchronously returns a GetPromptResult containing an optional human-readable description and a required messages array of PromptMessage objects. Each PromptMessage carries a role (user or assistant) and a content block that can be TextContent, ImageContent, or EmbeddedResource. The client injects these messages directly into the conversation context before the next model turn. \n\nArgument validation is the server's responsibility. If the caller omits a required argument or passes an argument of the wrong type, the server should return -32602 InvalidParams with a descriptive message field — not silently inject a partially-rendered template. Clients should treat a missing or empty messages array as a soft error and avoid injecting blank turns into the context, which can confuse the model. \n\nTiming matters: prompts/get is a synchronous request-response pair with no streaming. For prompts that involve expensive template rendering (e.g., fetching live data to embed), the server should complete the full render before replying rather than streaming partial results. If the prompt list changed between a prompts/list call and a prompts/get call — signalled by notifications/prompts/list_changed — the server may return -32602 if the named prompt no longer exists. Clients should be prepared to re-list and re-present choices to the user in that scenario.
When to use
When the user clicks/types a prompt (slash command).
When NOT to use
For model-driven workflows — use tools/call.
Notes
Capability gate is mandatory
The server must include a non-null prompts key in its capabilities object during initialize. If it does not, the client must not send prompts/get and should surface a configuration error rather than attempting the call and handling a MethodNotFound response at runtime.
Error codes to handle explicitly
Return -32602 (InvalidParams) when a required argument is missing or the named prompt no longer exists after a list change. Return -32601 (MethodNotFound) only if the capability was never advertised. Clients should distinguish these two codes: -32602 can be surfaced to the user as a fixable input error, while -32601 indicates a server configuration mismatch that the user cannot resolve at the argument level.
No pagination — messages array is atomic
Unlike prompts/list, prompts/get has no cursor or pagination mechanism. The entire PromptMessage[] is returned in a single response. If a prompt template would produce an extremely large context (e.g., embedding a full document), the server should truncate or summarize content internally rather than returning a payload that exceeds the client's context window limits.
Race condition after list_changed
Between receiving a notifications/prompts/list_changed notification and completing a new prompts/list round-trip, any in-flight or queued prompts/get calls that reference a stale prompt name may receive -32602. Design your client to queue a re-list and retry rather than surfacing this as a terminal error to the user.
Arguments map is strings only
The spec defines arguments as Record<string, string> — every value must be serialized to a string before sending. SDKs like the TypeScript MCP SDK enforce this at the type level, but Python implementations may silently coerce integers or booleans. Explicitly stringify all argument values on the client side to avoid subtle cross-SDK interoperability failures when servers perform strict type checks on received arguments.
Request parameters
| Name | Type | Purpose |
|---|---|---|
| name | string | Prompt name. |
| arguments | object? | Map of argument name → string value. |
Response fields
| Name | Type | Purpose |
|---|---|---|
| description | string? | Human description. |
| messages | PromptMessage[] | Messages to inject. |
Examples
Get a prompt
{ "method": "prompts/get", "params": { "name": "plan_vacation", "arguments": { "destination": "Barcelona" } } }Common mistakes
❌ Returning a single string
✅ Always return messages: PromptMessage[].