Run · ModulePlanned (Phase 7)
Leash
Hosted AI agents whose keys sign only one-time, action-bound approvals. A copied key cannot act.
The problem it prevents
A stolen key or a rogue operator quietly pushing a backdoor or draining a treasury.
Who it is for
Agencies running many sites; crypto teams with a treasury.
How it works
- 01Create an agent: patching, deploy or treasury. Its keys live in secure hardware and never export.
- 02Set an approval policy.
- 03Each proposed action produces one pending approval, bound to that exact action.
- 04Approve or deny. Every executed action gets a receipt on the public log.
What you get
- An agent list with status and targets.
- Pending approvals, one per exact action.
- Action history with receipts.
- Pause or revoke at any time.
Limits and caveats
- Phase 7; shape may change.
- Alert channels and team roles are not decided.
Questions
What stops a stolen key?
The key only signs approvals tied to one specific action. Without a fresh approval, it cannot act.
Who runs the agents?
ByteGuard hosts them; later the Network (Phase 8) can.
Which actions?
Patching servers, deploys and treasury moves first.
Is there a receipt for every action?
Yes. That is the "done when" for Phase 7.