The situation
Where they started.
Staff wanted to ask the core system questions in plain language: account status, transaction history, exception queues. The institution's IT security team, reasonably, would not allow an assistant that could act on the core system on its own, and would not allow the core's data to leave the tenancy.
What FlatClaw does
What was built.
- MCP connectors that mirror the core platform's real request shapes rather than inventing an abstraction over them.
- Read paths open to the roles that need them, with per-user credentials so an answer never exceeds what that person could see themselves.
- Write paths composed by the agent and paused in the approval queue; on approve, the request replays with the approver's own credentials.
- Every decision and every replayed request recorded in the audit log.
Results
What changed.
- A useful assistant that cannot act alone.
- Security review passed on the strength of the approval engine and the audit trail.
- Consistent answers for staff who previously escalated to a specialist.
- A pattern the institution can extend to other systems.