When an AI answer is wrong, “check the logs” is not enough.
Which logs?
- the chat log
- the MCP tool call
- the SQL query
- the permission decision
- the result contract
- the final answer
Production MCP database servers need observability that connects those steps.
At minimum, I’d want every database-backed tool call to record:
- user/workspace/tenant scope
- tool name and schema version
- parameters
- approved view or table surface
- database role
- duration
- rows scanned and returned
- timeout/truncation state
- policy decision
- freshness
- answer provenance
Otherwise you can see that a query ran, but not why the agent chose it or whether the answer was safe to trust.
Longer version: Observability for MCP database servers
SQL logs tell you what happened. MCP observability should tell you how the answer happened.
Top comments (0)