Safety from structure, not supervision.
Aggressive autonomy is only trustworthy if the safety is structural. Orcheo's guarantees are built into budgets, isolation, and hard gates — enforced by the system itself, never by someone watching it work.
You bring your own keys. On the Managed plan the model keys live in the platform's own vault; on the BYOK plans every credential sits in your per-tenant vault. Either way, the high-risk accounts — your cloud, your payment rails, your identity — stay owned by you.
Secrets are injected at runtime from the vault; the agent never sees the value.
Every table in the control plane is scoped by tenant and enforced with Postgres row-level security. A goal can only ever see the data of the tenant that started it — isolation is a database invariant, not a code convention.
Every goal carries a hard budget cap. It bounds what a run can spend, and with it the blast radius of a mistake or an abuse. When a goal reaches its ceiling it stops and asks rather than spending past it.
Regulated actions — cold email, automated calls, payments — enforce their rules inside the tool itself. Consent, throttling, and disclosure are part of the call, not a checklist bolted on afterward, and there is no path that bypasses them.
Side-effecting calls carry idempotency keys, so a retry cannot double-charge or double-send. Execution is sandboxed. Every action is recorded. Safety comes from the shape of the system, not from a human watching it run.
Questions? security@orcheo.ai