Savth

AI agents can propose trades, payments, subscriptions, and treasury operations, but giving a probabilistic system an unrestricted signing key turns one bad prompt, compromised tool, or hallucinated address into an irreversible loss. Savth separates intent from authority. An agent may request an action, but a deterministic policy layer decides whether it is allowed, requires human approval, or must be rejected.

Savth creates an expiring delegated authority with explicit program and asset allowlists, per-transaction and rolling spend ceilings, rate limits, destination constraints, and emergency revocation. Every proposed transaction is decoded, simulated, and compared with its stated intent before signing. The owner receives a plain-language approval request for anything outside the safe envelope, while an append-only receipt records the policy, simulation, and decision without exposing secret material.