Loops Housenear-agent-guard

NEAR HACKATHON
Links
Team
1 member- CHOwner
Chris
Overview
Problem: an AI agent that pays invoices reads text written by strangers. Give it an ordinary wallet key and one fake invoice becomes a real payment; telling the model to be careful does not bound the loss.
What we built: near-agent-guard, an agent wallet whose limits sit where the model cannot reach them. A Rust vault contract on NEAR testnet enforces per-agent per-transaction and per-window caps, a receiver allowlist, a per-window cap on ledger entries and revocation, and records every decision, approved or denied, in an on-chain ledger. The agent's only key is a NEAR function-call access key restricted to the vault's request_spend method, so NEAR itself rejects attempts to raise its own limits, attach deposits or transfer directly. The agent (mini-swe-agent with DeepSeek V4.1 Flash) runs in a no-network container with no keys; its requests are parsed inside a WasmBox WebAssembly sandbox, and every attempt is signed into a hash-chained audit log whose ids are stored on-chain, so log and chain verify against each other; an edited or cut-off log is caught.
Measured: with a real invoice and one from a lookalike supplier, the agent tried to pay the fake one in 20 of 20 runs; the vault refused all 24 attempts. The vault is deployed, locked and verifiable from the README with no keys; a scripted path needs no LLM key.
How NEAR fits: function-call keys and contract-held funds are the enforcement layer, and the on-chain ledger is the record that survives a compromised host. WasmBox is our pre-hackathon baseline; everything else was built in the window.
Demo post: we do not use or have an X/Twitter account, so the demo video is posted on Nostr instead (the demo link); the post says it was built at NEAR HACKATHON, carries #NEARHackathon and labels the AI voiceover.