Loops HouseAgentGuard

NEAR HACKATHON
Links
Team
1 member- COOwner
Code Master
Overview
AI agents can now pay for things, and a single injected email, web page or tool result can make an agent send money to an attacker. Giving an agent a wallet key means trusting every piece of text it reads.
AgentGuard is a spending firewall for AI agents on NEAR. Funds stay in a vault inside a NEAR smart contract, and the agent holds only a function-call key that can ask to spend. The contract decides every request against the owner's policy: per-payment caps, period budgets, rate limits, allow and deny lists, and approval for new recipients or large amounts. Every request ends paid, held for the owner, or refused with a reason code.
A prompt-injection risk oracle runs inside Intel TDX on OutLayer, reads the text that led the agent to pay, and signs a verdict the contract verifies on-chain. Agents connect over MCP (Claude Code, Codex or any client), and cross-chain payments go through NEAR Intents with each quote checked first. Owners get a dashboard with approvals, a kill switch, an activity log and alerts. Live on NEAR testnet, with a public demo vault that needs no wallet.
NEAR is what makes it possible to separate power from money: a contract-scoped vault, function-call keys, and a TEE oracle that signs what the chain can verify.