Loops House.Leash

NEAR HACKATHON
Links
Team
1 member- FROwner
Frank Olien
Overview
The problem: letting an AI agent pay for things today means handing it a private key or a card. A key signs whatever it is told, and the only thing between it and your money is the prompt, which a web page or a tool result can rewrite.
Leash is a spending account for AI agents. The owner signs in with Google, Apple or email, funds a USDC vault in the Leash contract, and gives each agent its own key with limits the agent cannot break: a daily cap, a per-payment cap, a list of who it may pay, a cooldown and an expiry. Inside the limits the agent pays on its own. Above them the payment parks and the owner approves or denies it on their iPhone with Face ID, from the Lock Screen if they want. An agent that needs more asks for a raise with a reason. Revoke deletes the key; pause stops every agent at once.
Before a first payment to a stranger, the agent gets signed facts about the recipient: account age, recent code changes, lookalike names, and a scam list shared across owners. Two reports and every agent on Leash refuses that account.
How NEAR fits in: the agent's key is a NEAR function-call key that can only call pay, request and ask_more on the Leash contract, so the protocol itself refuses any attempt to move tokens, add keys or change its own policy, before any contract code runs. The contract then enforces the owner's limits on chain, and every payment, refusal, approval and raise is a transaction anyone can read.
It works with Claude Code, Codex, Cursor and ChatGPT through an MCP server on npm, and with the Claude app through a hosted connector that shows a payment card in the chat. Agents can also pay APIs per request with x402.
The attack playground at leash.frankolien.com/attack runs a real model against a real vault, live: a poisoned invoice plus ten raw attacks, all stopped on chain, $0.00 lost. Open source, on NEAR testnet, with a 60 second check in the README that the AI judges can run.