ABAP Model Context Protocol Server SDK v2 (mcp2)

A stateless, rewrite of the ABAP Model Context Protocol server SDK, targeting the upcoming 2026-07-28 protocol generation while staying downward compatible with the legacy era (2025-03-26 … 2025-11-25).

MCP 2026-07-28 draft synchronous request/response profile. SSE, subscriptions, streaming responses, sessions, and server-pushed notifications are unsupported. Legacy revisions are served statelessly with the latest legacy data shapes.

This is a server implementation only. It is stateless — no protocol sessions, no SSE — so notification- and subscription-based features are out of scope by design (see docs/ProtocolSupport.md).

Status: 0.1.0 — first beta (changelog). Pre-release: the ABAP API surface is not frozen and breaking changes may land in any release before 1.0.0. Active v2 implementation for the stateless 2026-07-28 protocol generation. Protocol core, data classes, HTTP runtime, config/factory, DDIC tables, demo servers and the request/response Tasks extension are implemented and covered by abaplint plus the external Jest integration suite in the sibling mcp2_tests repository. The original session-based SDK lives separately and is unaffected; this package uses the distinct mcp2 object prefix so both can coexist in one ABAP system.

Main Differences from V1 MCP SDK

• Currently no 7.x downport - planned to be added with release, like within 1 month of spec release
• No sessions, etc.
• No default auth object and check delivered --> you must implement your own auth checks if required, see Configuration and security
• No table maintenance due to frequent install issues

Intentionally not supported

• ABAP Cloud - with SAP pushing towards using their MCP strategy and features in BTP I see no reason to put in the required effort

Documentation

User guide — building servers with the SDK:

• Getting started — install, ICF setup, first server
• Learning path — staged exercises from first tool to MRTR and tasks
• Demo servers — feature map and legacy/modern walkthrough
• Server identity and request context — metadata, capabilities, tracing, caching
• Tools — arguments, content blocks, structured output, validation
• Resources — resources, reads, templates
• Prompts — prompt templates and arguments
• Completions — argument autocompletion
• Schemas — schema builders (general / DDIC / elicitation) + validator
• Input required (MRTR) — elicitation, sampling, URL mode
• Tasks — long-running background work
• Configuration and security — tables, CORS, auth, error mapping
• Troubleshooting — symptom-first index of common failures
• Migrating from v1 — what changes, what is dropped, suggested order

Reference — design and internals:

• Class reference — every public class, what it is for, where it is explained
• Architecture — design, packages, eras, request flow
• Protocol support — versions, methods, what is excluded and why
• Conversion — per-field data-class conversion conventions
• Platform abstraction — HTTP/DDIC wrapping and the ABAP Cloud roadmap

Design at a glance

• Stateless only — every request is self-contained; no Mcp-Session-Id, no session tables.
• One server contract, two protocol eras — legacy initialize and modern per-request _meta negotiation served by the same zif_mcp2_server.
• Typed authoring surface — common server-authoring paths need no hand-written JSON: typed request getters (require_arg_string, get_arg_string_or, bind_arguments), typed response factories (zcl_mcp2_resp_call_tool=>text, =>error_text, =>structured_data) and full-control builders (add_resource, add_resource_link, set_structured_data) with typed content annotations (zif_mcp2_content), fluent schema builders (zcl_mcp2_schema_builder for tool schemas, zcl_mcp2_elicit_schema for flat elicitation forms), and typed input-request builders (zcl_mcp2_input_elicitation, zcl_mcp2_input_sampling) unified by zif_mcp2_input_request. UI icons are typed everywhere the spec allows them: serverInfo (get_icons), tools, prompts, resources, resource templates and resource links (zif_mcp2_content=>icons). Pre-built ajson remains an escape hatch for rich prompt content and custom extensions.
• Tool catalog base — inherit from zcl_mcp2_tool_server_base, declare define_tools( ) once, and implement call_tool( ). The base derives supports_tools, tools/list, get_tool_schema, unknown-tool rejection, and default schema validation from the catalog. Advertised-schema violations are returned as isError tool results in both eras; structurally malformed calls and unknown tools remain -32602.
• Era-aware capability checks on the base class — client_supports_elicitation, client_supports_elicit_url, client_supports_sampling, client_supports_tasks (modern: declared extension; legacy: per-request params.task opt-in), era_is_modern, get_protocol_version, plus can_request_input / input_required for MRTR, let authors safely gate era- and client-dependent features instead of hitting a framework -32021/-32600. Of the features SEP-2577 deprecates in 2026-07-28, Roots and Logging are not implemented; Sampling is kept (as an MRTR input builder) because it is the only key-less LLM access path from ABAP.
• Request context without _meta parsing — client identity, negotiated protocol, log-level hints and W3C trace context are available through typed base-class getters.
• serverInfo is written twice in the modern era — deliberately. The 2026-07-28 draft moved serverInfo out of the top-level result and into _meta["io.modelcontextprotocol/serverInfo"] (the ResultMetaObject shape) on 2026-07-16, so the SDK emits it there on every modern result. It also still writes the superseded top-level field on server/discover, because every published TypeScript v2 SDK release (through 2.0.0-beta.4) bundles a DiscoverResultSchema that requires it — omit it and the SDK's own client.connect() version-negotiation probe misclassifies the server as non-modern, silently downgrading real clients to the legacy era. Writing both costs a few bytes and keeps the SDK correct against the spec and usable by the clients that exist today. The top-level write is scheduled for removal once a released SDK reads _meta instead (details). Legacy initialize is unaffected: serverInfo stays a plain top-level field there.
• Conversion in the data class — explicit per-field JSON↔ABAP; central code only for the JSON-RPC envelope and the modern result stamp.

Target / tooling

• ABAP 7.5x classic (first iteration). abaplint v752.
• ABAP 7.4x donwport planned to be added after beta phase, cloud-ready not planned.
• Distributed via abapGit.

Used ABAP open-source projects

• ajson — JSON library, vendored as zmcp2_ajson (MIT).
• abaplint — linting and transpiler-based unit tests.
• abapGit — source control.

More awesome ABAP projects at dotabap.org.

Development checks

• Static check: npm install once, then npm run lint (abaplint, config in abaplint.json).
• Behavior: the Jest integration suite in the sibling mcp2_tests repository, run against an ABAP system serving the SDK and its test servers. It covers both protocol eras against the reference TypeScript SDK plus raw-HTTP wire checks.
• ABAP unit tests run in the ABAP system (SE80/ADT); the transpiler-based local run only covers a subset (no database, vendored-ajson unicode test excluded).

License

MIT. Vendored dependencies retain their original licenses; see THIRD_PARTY_NOTICES.md.