Keelward sits between your coding agent and Postgres. Normal queries pass straight through. Destructive ones wait for your approval, and every risky change can be undone.
Free for one database during early access.
Your agent wants to delete the customers table (4,812 rows). Its task: “clean up test data before the migration.”
SELECT queries and other read-only work run normally. Your agent never notices Keelward is there.
Targeted INSERT and UPDATE queries run, with a before-copy of every row they touch.
DROP, TRUNCATE, DELETE without a WHERE, and bulk changes over your limit pause until you approve them. No answer means no change.
Keelward keeps the real credentials. Your agent never sees them.
Give your agent the Keelward address instead of the database's.
Start with safe defaults, then decide what needs approval.
Approve or block from your phone or Slack when something needs you.
Before any risky change, Keelward saves a snapshot of the affected data. If an approved change turns out wrong, restore it from the history page.
Not if Keelward holds the only credentials. The agent can only reach your database through it, whatever tool or script it uses.
No. Only destructive actions ask you. Everything else passes or is logged, and you set the limits.
Keelward is built to add as little delay as possible. We'll publish real measurements before launch.
Not yet. We're building it now and inviting early users in small groups.