The key lives in the chip.
It is generated inside the security chip most computers already have, and it never leaves. The chip signs; the agent only asks. There is no key in memory and none on disk.
Agent keys
The key never leaves the chip, and the blockchain itself enforces the owner’s spending limits.
Public testnet
An AI agent that pays for things usually keeps its wallet key in a file. Whoever can read that file owns the money: malware, a leaked backup, a poisoned dependency, a tool the agent was tricked into running.
Agent keys removes the file. The agent can still pay by itself, but only through a chip that will not give up the key, and only inside limits its owner set.
It is generated inside the security chip most computers already have, and it never leaves. The chip signs; the agent only asks. There is no key in memory and none on disk.
The agent acts through an account that checks each action against the owner’s limits: where it may pay, how much each time, a total budget and an end date. Anything outside them is refused by the chain itself, whatever the agent, or whoever controls it, asks for.
The owner’s key is separate and never on the agent’s machine. It sets the limits, and replaces or revokes the agent’s key, each in one step. A cold recovery key can take over if the owner’s device is lost.
It works with several open-source agent frameworks in place of their key file, and it can pay for APIs per request.
| What happens | A key in a file | Agent keys |
|---|---|---|
| Malware reads the disk | All funds gone | Nothing. There is no key file. |
| A prompt injection says “send everything” | Sent | Refused by the chain. |
| Malware takes over the agent itself | All funds gone | At most what the limits allow, until the owner revokes the agent. |
| The laptop is stolen | The key is copied | The key can’t be extracted from the chip. The limits still apply, and the owner revokes. |
The key can’t be copied off the computer, on any payment rail. Malware on that computer can still ask the chip to sign while it is there, which is why the limits matter.
There the chain refuses anything outside them. On other rails the key is still held in the chip, and the limits have to be enforced by whoever runs the rail.
On a public testnet, with keys in the security chips of Linux and Windows PCs and of Macs, and in an open-hardware secure element.
It is for teams that build agent products, wallets and payment tools, and for people who run agents with real budgets.