Leashh
Agents fix prod, pay over x402 on Hedera, and can't spend a cent past their ENS limits. That's it.
Demo
About
Leash solves a massive problem with AI agents: how to let them spend money without letting them empty your wallet. Right now, if an agent needs to pay for a server or an API, you hand over a private key, and it has unlimited access. With Leash, every AI agent gets an ENS name with strict, on-chain spending rules. You can definitively state: “This agent has $5, can only buy AI services, and expires in 30 days.” **How It Works:** * **The Family Tree:** Agents form a hierarchy (e.g., parent -> child). A parent can give a child a slice of its budget, but the child *never* gets more money or a longer timeline than the parent. Smart contracts enforce this. * **The 6-Step Check:** Before *any* payment happens, Leash checks the agent (and its parents) for: name existence, remaining budget, per-request limits, allowed services, allowed currencies, and expiry. If one check fails, the payment is blocked. No money moves. * **The Kill Switch:** One on-chain transaction revokes an agent. Every child under it stops instantly, and rented servers shut down within 20 seconds. **What Agents Can Actually Do:** * **Buy AI:** Agents pay per token (via x402) using USDC on Hedera—no subscriptions or API keys needed. * **Rent Servers:** Agents can request infrastructure (like a Redis cache) and pay per minute. * **Auto-Fix Apps:** A team of agents can monitor a web app, diagnose issues, and pay for simple fixes (like restarting a service) within their limits. Every single action is logged publicly on Hedera. It's built on ENSv2 (using Ethereum Sepolia) for permissions, proving agents can be useful *and* financially constrained.
How it's made
Leash is an AI agent framework built on a simple split: the rules live on ENS (Sepolia), and the money moves on Hedera. The Rules: Every agent gets an ENSv2 subname. We store their limits—like budgets and allowed services—as simple text records. Because of how ENSv2 works, revoking a parent name instantly shuts down all its child agents. We even restricted transfers so agents can't go rogue and sell their names. The Money: Our Rust backend uses the x402 protocol to settle micro-transactions in USDC on Hedera. Hedera’s tiny fees make paying a fraction of a cent per AI request actually viable. Before any money moves, a policy engine verifies the agent's ENS rules, and it only pays for successful LLM responses. Look, Ma, No Database: We ditched traditional databases entirely. Instead, Hedera Consensus Service (HCS) acts as our audit log, service directory, and cross-provider spending ledger. What it does: Agents can rent Docker containers by the minute or fix broken apps. In our "autopilot" setup, three agents team up: one monitors the system, one pays an LLM (like Llama or Qwen) to diagnose issues, and the third pays to execute a fix from a strict, pre-approved menu. The Setup: Our React frontend reads directly from Sepolia and Hedera mirror nodes, so it works even if our own backend goes down. The Coolest Hack: If you revoke an agent's name on Sepolia, our cleanup loop catches it and literally kills that agent's Docker containers on a completely different server within 20 seconds. It’s real-time, cross-chain mandate enforcement.