Loops HouseAgentGuard

NEAR HACKATHON
Team
1 member- WEOwner
WEI WU
Overview
Problem: AI agents increasingly hold wallets to pay for APIs, compute and data, and a single prompt injection can drain them because the agent's key can sign anything. AgentGuard is a policy gateway between an AI agent and its NEAR wallet. The agent never holds the signing key: it submits transaction requests to the gateway, where a Python policy engine enforces per-tx and rolling 24h limits, a hard cap, recipient allowlists, contract/method whitelists, blocked methods and actions (approve, ft_transfer_call, AddKey, DeployContract, DeleteAccount...) and rate limits. Each request gets ALLOW, REQUIRE_APPROVAL or DENY with plain-language reasons; over-limit requests wait in a human approval queue (CLI + FastAPI panel) and everything goes to an append-only audit log. NEAR's role: if the agent needs its own key, AgentGuard provisions a NEAR Function-Call access key restricted to one receiver, specific methods and a gas allowance, so even a fully compromised agent is bounded by the protocol itself. Testnet demo: an agent paying a vendor is signed; a prompt-injected agent trying to send 3 NEAR is blocked and logged.