Loops HouseKey proof

NEAR HACKATHON
Links
Team
1 member- BEOwner
Benny
Overview
The problem: people give AI agents wallets and write the limits in a prompt, like "never go past 5 NEAR". A prompt is only text. What an agent can actually do is decided by the permissions on its account's access keys, and nobody checks that the two agree.
What I built: KeyProof, a read-only checker. You write the spending policy in plain English from a fixed template (an allowance limit, a contract allowlist, no full access keys, named methods only). KeyProof reads the agent account's keys and returns a verdict per promise: holds, can be exceeded or unknown, plus the exact key field that decides it, for example keys[1].allowance = 12 NEAR against a 5 NEAR limit. It also flags full access keys, keys with no allowance (unlimited gas), keys that can call any method, and receivers that are not on the allowlist. It never asks for a private key and never signs anything. It checks key permissions, not whether the policy itself is sensible.
How NEAR fits in: NEAR's function call keys let you scope an agent to one contract, named methods and a gas allowance, and those fields are readable on chain with the view_access_key_list RPC call. KeyProof turns them into a plain-English verdict. Testnet only. Built with TypeScript and Vite: a landing page, a checker with a live testnet mode and recorded before and after snapshots, and a text report.