An AI answer from a live database should not end at the sentence.
If an assistant says:
“MRR is down 8% in EMEA.”
The next question is obvious:
Based on what?
For MCP database servers, citations should be part of the answer contract:
- query ID
- approved view or table family
- timestamp and freshness window
- row count or sample status
- filters applied
- redaction status
- metric definition used
- audit event ID
“Source: database” is not enough.
The answer needs a source trail that a human can inspect when the number matters.
Longer version: MCP database answer citations
Natural-language summaries are useful. Evidence is what makes them safe to trust.
Top comments (0)