Artificial Intelligence Blogs Posts
cancel
Showing results for 
Search instead for 
Did you mean: 

In my previous blog post I described installing an MCP Server which wraps ABAP Development Tools API, and extending it with some WRITE functionality - all conversationally with AI Agents.

I asked Perplexity to go look into GitHub, see what other MCP Servers people have built. She recommended mcp-abap-adt from fr0ster as the best all-rounder: https://github.com/fr0ster/mcp-abap-adt.

k1000_6-1772895572947.png

I again used an Agent in Google Antigravity to clone the repo and install it. Updated *.env configuration manually and asked the Antigravity Agent to update the configuration for OpenAI Codex, that I had installed as a desktop app on my Mac. That didn't work flawlessly and Codex (GPT-5.3) had to go through some shenanigans to make it work, but it the end he managed. He queried the MCP and found that it exposed massive 131 tools:

k1000_1-1772894619783.png

Nice. Good job fr0ster and team!

In the spirit of vibe-coding, I asked Codex to create a simple SALV Report. That took about 84 seconds:

k1000_2-1772894721801.png

k1000_3-1772894744425.png

Codex proactively suggested that we should add a Selection Screen. I specified the requirement, and he added the Selection Screen:

k1000_4-1772894933484.png

k1000_5-1772894943130.png

What he couldn't add were Text Elements for the Selection Screen Fields. No tool for that in the MCP - ADT API itself probably doesn't expose such functionality.

Next I inquired whether he would be able to build a full-stack Fiori app. He confidently replied "Why yes of course, I thought you'd never ask" (OK, he didn't reply exactly that, but something in that spirit). I specified an app and he proceeded to make a pig's ear of it. Probably my fault - I suspect my specification was poor. It's all about the Specification in the AI age, people! The better the Prompt, the better the result. Otherwise garbage in - garbage out.

Oh, and Context. Context Engineering is king.

And Intent for more complex tasks. But that's it, I swear.

 

So there you have it - a bit of vibe-coding, ABAP style. Joule it ain't, but beggars-choosers, I guess. Joule will have been trained on a much larger ABAP codebase. Claude, Codex and Co. can only see what's in GitHub, maybe SAP Help,  and various publicly-available documents.

Time permitting, I'll get back to the Fiori App topic (also with Claude). I have a feeling there is more to be achieved there.

 

 

 

2 Comments
Labels in this area