Loops HouseSecondEyes

NEAR HACKATHON
Links
Team
1 member- PROwner
praise
Overview
People are starting to give AI agents money, and an agent can be tricked. A fake invoice saying a supplier changed accounts can make it pay a scammer, and spending limits do not help because the scam is under the limit.
SecondEyes lets you give an agent a budget and a plain-language instruction. Every payment is checked against that instruction by an independent model on NEAR AI Cloud, and only an exact, signed ALLOW releases the money. A live demo shows a real model falling for a scam invoice, with the payment refused and the money staying put. Any agent can connect through MCP or the command line.
NEAR is what makes it enforceable. The agent holds a function-call key that can only ask to pay. The contract rebuilds each request itself and verifies the judge's ed25519 signature on-chain, so neither the agent nor my server can fake a yes. The contract is locked, with no full-access keys.
On 47 test cases I wrote after freezing the judge's wording, run three times each, it paid 0 of 90 attacks and 51 of 51 honest payments, against 69 of 90 attacks for limits alone. The README also lists what it cannot catch. Runs on testnet with test money.