Control and trust

Put a supervisor between the model and the money

A daemon that prepares, explains, signs and tracks every transaction your agent sends.

The agent daemon runs beside your agent. The model asks for an action; the daemon shows what it would do in plain terms, applies your policies, signs, submits and follows the transaction through to final.

Install the CLI Developer docs

A language model is good at deciding what to do and bad at being trusted with a private key. The agent daemon splits those jobs. Your agent asks for a payment or an order. The daemon builds the transaction, decodes exactly what it will do, checks that against policies you wrote, and only then signs and sends. The model never touches the key.

What the daemon takes off your plate

Prepares each transaction and explains it in plain terms before it is signed, so a person or a policy can judge it
Applies your policies, from a hard spending cap to a rule that any new counterparty needs a person's approval
Holds approvals that bind to the exact transaction approved, so what was shown is what gets sent
Tracks every transaction from queued to final and resolves anything uncertain from the network's own records
Recovers safely after a crash or restart. The same transaction is resent byte for byte, never duplicated

An afternoon at a growing startup

01
The agent proposes
A support agent decides to refund 120 USDC to a customer.
02
The daemon checks
Policy allows refunds up to 100 USDC without a person. This one is above the line, so the daemon holds it and sends it to the founder's wallet.
03
A person approves
The founder sees the exact refund, the recipient and the reason, approves in one tap, and the daemon signs and submits.
04
It is tracked to final
The receipt returns in milliseconds and the daemon records the proof. Had the laptop died midway, the daemon would have resumed on restart without a second refund.

Policies, not prayers

Give the model good tools and a clear policy, and keep the signature where you can see it.

Teams usually start with one rule: any transaction above a threshold needs a tap. From there they add limits per counterparty, quiet hours and a short allowlist of services. The policies live with the daemon, so a change takes effect at once without redeploying the agent.

Works with your assistant

The MCP server and the A2A transport route every write through the daemon, so a desktop assistant or another agent gets exactly the authority you set and nothing more.