Loops HouseAgentTrust

NEAR HACKATHON
Links
Team
1 member- MOOwner
Mohit Kumar
Overview
Problem: an AI agent with a wallet can pay anyone. One bad prompt or one compromised tool can drain it, and a limit written in the prompt is only a suggestion.
Product: AgentTrust puts the owner's rules in the payment path. The owner sets, once, which exact supplier identities are trusted, which capability credential a supplier must hold from an issuer the owner accepts, and a cap per purchase. The agent gets one payment tool over MCP.
How NEAR fits: the rules live in a native NEAR contract. The agent holds a NEAR function-call access key that can call one method, execute_payment. It cannot change the rules or move funds any other way. The contract checks the supplier's identity, credential, limits and signed quote before it reserves and transfers NEAR, so skipping our SDK does not skip the rules: a direct call with a missing credential was rejected on-chain with no payout.
What ran on NEAR testnet: a local model (qwen3:4b through Ollama and MCP) was asked to get an invoice file audited. Supplier A lacked the credential and was refused before any transaction was sent. Supplier B was approved, 0.002 NEAR settled, and only then did B release its report (15 test invoices, 7 rejected), which the buyer checked separately. An exact retry of a settled payment paid nothing twice.
Limits: testnet only, synthetic invoices, and we ran both suppliers and the credential issuer ourselves. The host offers the model one allowed supplier at a time and supplies the payment ID and the file. One native refund test case is still open. The website and docs design are reused from earlier AgentTrust work; the NEAR contract, SDK, paid HTTP flow and MCP server were built in the hackathon window.
Site: https://www.agenttrust.tech Docs: https://docs.agenttrust.tech