SAP ABAP MCP Server SDK

Dynamically creates MCP server instances at runtime using the ZCL_MCP_SERVER_FACTORY class, which reads server metadata from ABAP configuration tables (ZMCP_*) and instantiates the appropriate server class via reflection. The factory pattern decouples server configuration from deployment, enabling zero-code-change server registration and multi-tenant server hosting within a single ABAP system.

The ZCL_MCP_JSONRPC class parses incoming HTTP request bodies as JSON-RPC 2.0 messages, validates structure (jsonrpc version, method, params, id), and routes method calls to the appropriate handler on the instantiated server. Implements full JSON-RPC 2.0 spec including error response formatting, batch request handling, and notification support (fire-and-forget calls with no id field).

Routes incoming HTTP requests to appropriate MCP servers based on ICF service path configuration. The ZCL_MCP_HTTP_HANDLER parses the request path, looks up the corresponding server in configuration tables, instantiates the server via the factory, and dispatches the request. Supports multiple servers per ICF service with path-based routing (e.g., /mcp/server1, /mcp/server2).

Provides utility classes for JSON serialization/deserialization, schema validation, and data type conversion within ABAP. Leverages ABAP's native JSON support (CALL TRANSFORMATION, /ui2/cl_json) to handle MCP protocol messages and custom data structures. Includes helpers for converting ABAP types to JSON-serializable formats and vice versa.

Provides development tools and demo server examples (zcl_mcp_demo_server_stateless) for testing MCP server implementations. Includes utilities for validating MCP protocol compliance, testing tool invocation, and debugging request/response flows. Demo servers demonstrate best practices for resource, tool, and prompt implementation.

Provides utilities (zmcp_clear_mcp_sessions program) for managing session lifecycle, including automatic cleanup of expired sessions and manual session termination. Prevents memory leaks from accumulated session state in long-running ABAP systems. Supports configurable session timeout and cleanup policies.

Supports multiple MCP specification versions (2025-03-28 and 2025-06-18) with version negotiation during the initialize handshake. Handles protocol evolution by validating client-requested capabilities against server-supported features, enabling forward/backward compatibility as the MCP spec evolves. Version information is exchanged during initialization to ensure client/server compatibility.

The ZCL_MCP_HTTP_HANDLER class validates incoming HTTP requests using ABAP's native authentication (user credentials, SSO tokens) and authorization (transaction codes, authorization objects). Integrates with SAP's ICF framework to extract user context and enforces ABAP-level access control before routing to MCP servers, enabling fine-grained permission control per server or per tool.

The ZCL_MCP_HTTP_HANDLER manages CORS headers (Access-Control-Allow-Origin, Access-Control-Allow-Methods, Access-Control-Allow-Headers) for HTTP requests from MCP clients, with policies configurable via ABAP configuration tables. Handles preflight OPTIONS requests and enforces origin whitelisting to prevent unauthorized cross-origin access to MCP servers.

Supports multiple session management strategies (stateful with server-side session storage, stateless with client-provided context) via configurable session handlers. Stateful mode stores session state in ABAP memory or database tables, enabling multi-request conversations; stateless mode validates client-provided tokens/context without server-side storage, reducing memory overhead. Session lifecycle is managed per MCP client connection with configurable timeout and cleanup.

Implements JSON Schema validation for tool parameters and resource outputs using ABAP's native JSON schema support. Schemas are defined in configuration or code and validated at runtime before tool execution and after resource retrieval, ensuring type safety and preventing malformed data from reaching business logic or being returned to MCP clients. Generates validation errors in JSON-RPC format when schemas are violated.

Enables ABAP servers to expose data as MCP resources (files, database records, API responses) through a standardized resource interface. Resources are registered with metadata (name, description, MIME type, URI) and retrieved on-demand by MCP clients. The framework handles resource discovery (resources/list), retrieval (resources/read), and template expansion for parameterized resources, allowing AI clients to access ABAP data without direct database access.

Exposes ABAP function modules and custom server methods as MCP tools that AI clients can invoke. Tools are registered with input/output schemas, descriptions, and execution handlers. The framework handles tool discovery (tools/list), invocation (tools/call), parameter marshaling, and error handling, allowing AI models to execute ABAP business logic (create orders, update materials, run reports) as part of agentic workflows.

Exposes ABAP-generated prompts as MCP prompts that guide AI model behavior. Prompts can be static templates or dynamically generated based on ABAP system context (current user, business data, system configuration). The framework handles prompt discovery (prompts/list), retrieval (prompts/get), and argument substitution, allowing AI clients to use ABAP-provided instructions and context to improve model responses.

Provides ZCL_MCP_SERVER_BASE abstract class that custom servers extend to implement MCP protocol handlers. Includes lifecycle hooks (initialize, shutdown, on_request) and abstract methods for resources, tools, and prompts. The framework handles protocol compliance, error handling, and session management, allowing developers to focus on business logic implementation without reimplementing MCP protocol details.