That’s a really nice design, Paul. Evidence plus a probability per insight, and “dreaming” consolidation alongside transactional updates, solves the freshness question I was asking about.
In the spirit of this thread, full disclosure: I’m the engineer behind SuiteBrain AI ( SuiteCRM inside ChatGPT, Claude, Gemini & Grok, Introducing SuiteBrain AI ), which is our attempt at the “SuiteCRM as system of record, intelligence alongside it” direction. Bastian’s point about generic API wrappers shaped a lot of it:
- Business-level tools, not just CRUD: convert a lead, close a case, move an opportunity stage, log a call, schedule a meeting
- It reads each instance’s own modules and fields at runtime, custom ones included
- Every user signs in as themselves, so SuiteCRM’s own ACLs apply to every call
- Write and delete tools are marked as such, so the AI app asks before acting
@BastianHammer, given you’ve tested several MCP layers, I’d value your honest view of where this one still falls short. Anyone curious can point it at a staging instance from suitebrainai.com.
And Paul, a knowledge layer like yours is exactly what a connector like this should read from first. I’d be glad to compare notes.