A human analyst may run five careful queries.
An AI workflow can run fifty tool calls while still looking like one user question.
That is why MCP database servers need rate limits that understand more than HTTP requests per minute.
The expensive unit might be:
- a tool call
- a generated query
- a result page
- a retry
- rows scanned
- rows returned
- a cross-source join
- repeated attempts at the same broad question
For production database access, limits should follow user, workspace, tenant, tool, data source, and query class.
And when the limit is hit, the tool should not just say “429.”
It should return a useful next step:
Narrow the date range.
Choose a tenant.
Use the approved revenue summary.
Ask for an aggregate before raw rows.
Longer version: Rate limits for MCP database servers
One helpful agent should not become the whole workload.
Top comments (0)