Kisenon is serverless Postgres with Agent-Safe Change Control (ASCC) built in: every agent works in a scoped, budgeted, fully-attributed sandbox — its changes are risk-linted, blast-radius-checked, and promoted to production only when they pass. And if something still gets through: one click, undone.
Two things make it safe: the agent physically can't exceed its bounds, and you can watch exactly what it did.
The agent works on a per-run fork with a scoped, non-superuser credential. It never holds a credential that can write main — promotion happens server-side, only for changes that passed in the sandbox. And every sandbox carries a budget — statements, compute, wall-clock: blow it and the sandbox auto-discards.
Every query the agent runs is attributed and streamed live to a read-only console. No summaries, no LLM in the loop — exactly what happened, as it happens. And every promote is a signed, offline-verifiable attestation: audit without trusting us.
Test migrations against live-shaped data, see the diff and the action log, then promote — self-serve for toy schemas, one-click human approval for prod, set per project. And every promote anchors an LSN first: if something slips through, one click undoes it.
Live today — sign up and point an agent at a live database.
Idle endpoints suspend after 5 minutes. Pay storage only — no per-hour compute floor. The OG noble-gas idea.
First connect after a fresh create provisions in 10–30s. After that, waking a suspended endpoint is sub-second same-region — ORMs never see a timeout.
Branches are pointers, not copies. Spin one up, run a destructive migration, throw it away. Free below the delta.