Give your AI agent a wallet, not a blank cheque.

AgentShield is a vault contract that holds an agent's funds and enforces the owner's rules on every payment. The AI can propose a payment. Only the vault decides whether it happens.

Set up an agent Try to rob our agent, no wallet needed Open the agent console

Your sentence→
AI model→
AgentShield checks→
Vault contract→
Paid or blocked

1 · Owner

Set the rules

Register an agent with its own signing key, set a per-payment and daily limit, approve who it may pay, and deposit MON into the vault.

Owner setup →

2 · Agent

Ask in plain language

An AI model such as OpenAI or Kimi turns “pay Shop 0.05 MON for coffee” into a structured request. The model only sees contact names, never addresses or keys.

Agent console →

3 · Vault

The contract decides

The vault checks the signer, recipient, limits, pause switch and replay protection onchain. It pays, or it reverts and nothing moves.

Address book →

Enforced by the contract, not the app

What the vault guarantees

Only approved recipients

Any other address is refused, even if the AI is tricked into naming it.

Per-payment and daily caps

Limits are checked by the contract on every payment, per UTC day.

Agent cannot withdraw

Its key can only call executePayment. Withdrawals go to the owner only.

Emergency stop

Pause blocks every payment instantly. Deactivate retires an agent.

Key rotation

If an agent key leaks, the owner swaps it in one transaction.

Verifiable receipts

Each request ID is a hash of the prompt, model, recipient and amount.

Deployed contracts

AgentRegistry: 0x7a91…d71a ↗
AgentGuardVault: 0xbdbb…a202 ↗

Honest limits

Testnet only. Amounts are test MON with no real value.
In this demo a person clicks Send as the agent. A production agent signs with its own key automatically, under the same vault rules.
The model name on a receipt is declared by the server, not signed by the model provider.