MCP Server SDK: Prompts

This documentation explains how to implement and use prompts in the Model Context Protocol (MCP) Server SDK. Prompts enable AI assistants to retrieve pre-defined conversation starters or templates from your ABAP systems.

Table of Contents

• Overview
• Prompt Request Classes
• Prompt Response Classes
• Implementing Prompt Handlers
• Examples

Overview

Prompts in the MCP protocol are pre-defined conversation templates that can be retrieved and used by AI assistants. They provide:

• Standardized ways to invoke specific capabilities
• Parameterized templates (with arguments)
• Multi-modal content (text, images, audio, resources)
• Consistent conversation patterns

Prompts are identified by a unique name and can accept arguments to customize their behavior.

Prompt Request Classes

List Prompts Request

The ZCL_MCP_REQ_LIST_PROMPTS class handles requests to list available prompts on the server.

Key methods:

• get_cursor(): Retrieves the pagination cursor if provided
• has_cursor(): Checks if a pagination cursor was provided

Get Prompt Request

The ZCL_MCP_REQ_GET_PROMPT class handles requests to retrieve a specific prompt with arguments.

Key methods:

• get_name(): Retrieves the requested prompt name
• has_arguments(): Checks if arguments were provided
• get_arguments(): Retrieves the arguments as a table of key-value pairs

Prompt Response Classes

Get Prompt Response

The ZCL_MCP_RESP_GET_PROMPT class builds the response for a prompt request, supporting various content types and annotations.

Key methods:

• set_description(): Sets the prompt description
• set_messages(): Sets complete message list
• set_meta(): Sets metadata for the prompt

Content methods:

• add_text_message(): Adds a simple text message
• add_image_message(): Adds an image message (Base64-encoded)
• add_audio_message(): Adds an audio message (Base64-encoded)
• add_text_resource_message(): Adds embedded text resource
• add_blob_resource_message(): Adds embedded binary resource
• add_resource_link_message(): Adds a resource link reference

Content Types:

1. Text Content: Simple text messages
2. Image Content: Base64-encoded images with MIME type
3. Audio Content: Base64-encoded audio with MIME type
4. Text Resource: Embedded text resources with URI
5. Blob Resource: Embedded binary resources with URI
6. Resource Links: References to external resources

Annotations Support:

All content types support optional annotations:

• audience: Target audience list
• priority: Priority level (decimal)
• last_modified: Timestamp of last modification

Prompt list entries and prompt arguments also support optional title fields. Prompt entries additionally support icons and meta metadata.

Implementing Prompt Handlers

Override these methods in your MCP server:

METHODS handle_list_prompts REDEFINITION.
METHODS handle_get_prompt REDEFINITION.

Handling List Prompts

METHOD handle_list_prompts.
  response-result->set_prompts(
    VALUE #(
      ( name = 'greeting'
        description = 'Greet a person by name'
        arguments = VALUE #(
          ( name = 'name'
            description = 'Name of person to greet'
            required = abap_true )
        )
      )
    )
  ).
ENDMETHOD.

Handling Get Prompt

METHOD handle_get_prompt.
  DATA(prompt_name) = request->get_name( ).
  
  CASE prompt_name.
    WHEN 'greeting'.
      IF request->has_arguments( ).
        DATA(arguments) = request->get_arguments( ).
        READ TABLE arguments INTO DATA(name_arg) WITH KEY key = 'name'.
        IF sy-subrc = 0.
          response-result->set_description( 'Greeting prompt' ).
          response-result->add_text_message(
            role = zif_mcp_types=>role_user
            text = |Hello { name_arg-value }, nice to meet you!|
          ).
        ENDIF.
      ENDIF.
      
    WHEN OTHERS.
      response-error-code = zcl_mcp_jsonrpc=>error_codes-invalid_params.
      response-error-message = |Prompt { prompt_name } not found|.
  ENDCASE.
ENDMETHOD.

Examples

Multi-Modal Prompt

METHOD handle_get_prompt.
  CASE request->get_name( ).
    WHEN 'analyze_media'.
      response-result->set_description( 'Analyze multiple media types' ).
      
      " Text instruction
      response-result->add_text_message(
        role = zif_mcp_types=>role_user
        text = 'Please analyze the following media files:'
      ).
      
      " Image content
      response-result->add_image_message(
        role = zif_mcp_types=>role_user
        data = get_base64_image( )
        mime_type = 'image/png'
        annotations = VALUE #( priority = '1.0' )
      ).
      
      " Audio content
      response-result->add_audio_message(
        role = zif_mcp_types=>role_user
        data = get_base64_audio( )
        mime_type = 'audio/mp3'
      ).
      
      " Resource link
      response-result->add_resource_link_message(
        role = zif_mcp_types=>role_user
        uri = 'https://example.com/document.pdf'
        name = 'Analysis Document'
        description = 'Supporting documentation'
        mime_type = 'application/pdf'
      ).
  ENDCASE.
ENDMETHOD.

Resource-Based Prompt

METHOD handle_get_prompt.
  CASE request->get_name( ).
    WHEN 'review_code'.
      DATA(arguments) = request->get_arguments( ).
      READ TABLE arguments INTO DATA(class_arg) WITH KEY key = 'class_name'.
      
      IF sy-subrc = 0.
        DATA(source_code) = get_class_source( class_arg-value ).
        
        response-result->set_description( 'ABAP code review prompt' ).
        
        " Add embedded text resource
        response-result->add_text_resource_message(
          role = zif_mcp_types=>role_user
          uri = |abap://classes/{ class_arg-value }|
          text = source_code
          mime_type = 'text/x-abap'
          annotations = VALUE #( 
            audience = VALUE #( ( 'developers' ) ( 'architects' ) )
            priority = '0.8'
          )
        ).
      ENDIF.
  ENDCASE.
ENDMETHOD.