Overview
A great deal of useful information lives in databases that no off-the-shelf AI tool can reach: the ERP, the order system, the internal platform. The usual shortcut — giving an AI a connection string — gives it far more than any single question needs.
Kroy connects to PostgreSQL with a dedicated role. Your team decides which tables, views and queries are exposed. Authorised AI asks Kroy, and Kroy runs only what policy allows.
Named queries first
The safest default is a set of named, parameterised queries written and reviewed by your team — open_orders_by_customer, stock_below_reorder_level. The AI chooses a query and supplies parameters; it does not write SQL against production. Ad hoc read-only queries can be enabled for particular people and agents, and require approval by default.
Permissions
Each query and statement is a Kroy capability. Kroy decides every request against the person, the agent, the delegation between them and the organisation’s policies. Kroy can also apply row-level rules — for example, a regional manager’s AI sees that region’s orders only.
Writes are named statements, such as marking an order as reviewed, and require approval by default.
Industries
Technology companies and firms with in-house systems can give AI useful access to operational data. In financial services, where data handling is closely supervised, named queries and a complete audit trail make it easier to explain exactly what AI could reach.
Security
Kroy holds the database credentials. AI clients never receive them. Every query is recorded with the person, the agent, the query name, the parameters and the number of rows returned, so a reviewer can see precisely what left the database.