Loops HouseAgent Treasury

NEAR HACKATHON
Links
Team
1 member- ISOwner
Iskander Kudashev
Overview
Problem: AI agents now do real work, but giving one money means giving it a private key, and a key can spend everything. One prompt injection and the wallet is drained. Approving every payment by hand kills the autonomy.
Agent Treasury is a bank for AI agents. The owner keeps USDC in a NEAR vault contract and gives each agent a mandate: per-payment and daily limits, an allowlist of recipients and chains, an expiry, pause and revoke. The agent never holds a key that can withdraw; it can only ask the vault to pay, and the contract checks the mandate and returns a machine-readable reason when it refuses. Any agent plugs in over MCP (Claude Code, Codex, your own).
How NEAR fits: the contract is the enforcement layer, so the rules hold even if the agent or its prompt is compromised. NEAR Intents (1Click) delivers each payout to the recipient's chain, so the agent never touches bridges. The vault is built as a Shade Agent contract (requires_tee on mainnet). NEAR AI Cloud turns plain text into rule proposals the owner signs.
Live on NEAR mainnet (vault.agenthack.near): Claude Code over MCP paid Maria in USDC on Base and Alex on Arbitrum through NEAR Intents; a 50 USDC request and a prompt-injected address were refused by the contract (PER_TX_LIMIT_EXCEEDED, RECIPIENT_NOT_ALLOWED). The vault funded a Hyperliquid account, where an AI agent traded HYPE through the same deterministic risk gate as rule-based bots; Stop cut it off instantly.
Reviewers: scripts/verify.sh runs all tests (contract, worker, sdk, mcp, dashboard) and checks the mainnet state read-only, no keys needed. Payroll is the first use case; the kit (contract, worker, MCP server, dashboard) is reusable for any agent that needs to spend.