ABAP Model Context Protocol Server SDK

June 22, 2025 ยท View on GitHub

Discord - for all ABAP-AI repos.

Overview

ABAP implementation of the Model Context Protocol based on version 2025-06-18.
Note that this is only a server implementation as I currently see no realistic way to support SSE streaming which is mandatory for the client.

Supported protocol versions: 2025-03-28, 2025-06-18

Note that old protocol versions might be dropped after a while especially if they increaase maintenance effort.

Documentation

See Overview in docs folder.

Used ABAP OpenSource projects

Special thanks to:

  • ajson integrated as zmcp_ajson
  • abaplint for review and more important 7.02 downport
  • abapGit for source code control

Find more awesome projects at dotabap.

Features & Limitations

Currently implemented:

  • Prompts
  • Resources
  • Tools

No SSE support --> no notifications and therefore no support of listChanged, subscribe and logging.\

Not implemented:

  • Sampling - considered for the future, open an issue if you have a use case that would benefit from it
  • Completions - questionable from a performance point of view with ABAP
  • Elicitation - without SSE I see no proper way to implement this
  • Roots - likely irrelevant for ABAP
  • Further notifications not relevant without SSE: Cancellation, Ping, Progress, Logging

Authorization

Implement the standard Authorization in ABAP via custom code is not feasible, we have to rely on ABAP features.
An overview of likely options is outlined in Authentication Documentation.

Autentication

OAuth is suggested but outside of newer S/4HANA releases challenging, for options see Authentication. Further suggestions are welcome.

Roadmap

See currently open issues for the next planned steps. Longer term I intend to add features to create modular MCP servers based on reusable implementations.

FAQ

See FAQ.

Contributing

Feel free to:

  • Open issues for bug reports
  • Use Discord for discussions
  • Suggest improvements
  • Pull requests are welcome but ensure they are tested and follow the overall style, if in doubt open an issue first

GenAI Notice

Part of the coding was generated using AI, mostly with Sonnet 3.7/4 and GitHub Copilot. Detailed documentation is largely AI generated - only high-level reviewed.