Weir Protocol

Demo

About

Your agent has a $12/month API subscription. You fund it with $500. In forty months it is dead, and so is whatever it was maintaining. Recurring costs do not stop. Budgets do. Every funded agent today is mortal by construction. You hand it a lump sum and a spending cap, and that cap is a policy: a session key that drains, an allowance someone can raise, a config line in a service you hope nobody edits. The agent spends principal because principal is the only thing it has. Weir endows the bill instead of funding the agent. An owner commits principal to Aave once. From then on the agent is paid only what the index produces, and three things are true by construction rather than by policy. The agent never chooses an amount: harvest() derives it from Aave's liquidity index as accrued = P x (i1 - i0) / i0, so there is no amount parameter to abuse. It never chooses a destination; that comes from storage. And it cannot reach the principal, because after withdrawing the accrual the remaining balance is P again. Because a caller has nothing to choose and nothing to gain, harvest is permissionless. No keeper, no relayer, no liveness dependency on us. The agent can call it and pay itself, and it has: 0.016172 USDC released on Base Sepolia with the principal unchanged. The owner can close the endowment and reclaim 100% of the principal at any time. At Aave's live mainnet USDC rate, $4,045 endows a $12/month subscription forever. The yield is not small. The bill is small, and the principal is simply whatever that bill requires

How it's made

An Aave aToken is a scaled-balance token: your balance is a raw balance times a global index. That is the same object as a Token-2022 scaled-UI multiplier on Solana, and it is the one property Weir needs. For principal P committed at index i0, the releasable amount at i1 is P x (i1 - i0) / i0. After withdrawing exactly that, the remaining balance is P again, so principal is preserved by construction rather than by a check someone could forget. Stack: Solidity 0.8.35 and Foundry for the contract; Next.js 15, viem and Privy for the app; a subgraph on The Graph's Subgraph Studio; an MCP server for the agent; a TypeScript agent with an x402 client. Two things were genuinely hacky, and they are the parts we are most pleased with. First, keeping the swap permissionless. harvest() has no keeper because the amount comes from the index and the recipient from storage, so a caller has nothing to abuse. Adding Uniswap reintroduced exactly one attackable parameter, amountOutMinimum, and a caller could pass zero and sandwich the agent. We floored it against Aave's own Chainlink-backed oracle, so a caller may RAISE the slippage guard but never lower it. That adds no trust assumption we did not already carry, since Weir already depends on Aave. Second, a rounding reserve. Aave burns scaled units with rayDiv round-half-up, so withdrawing the exact accrual burns up to one extra scaled unit and leaves the balance a hair BELOW principal. Left alone that leaks a unit per harvest and, with several endowments sharing one aToken position, the shortfall compounds until the last owner to close is short. Weir withholds index/RAY + 1 units per harvest, tracks the total, and pays it to the agent on close, so nothing leaks and nothing is stranded. A fork test runs 48 harvests across two endowments and asserts both. Partner tech, and how it earned its place. The Graph indexes every endowment and release, and our MCP server answers the question an endowed agent actually needs — not "what is my balance", which is actively misleading here because principal is visible but unspendable, but "what can I spend forever". It returns a PERPETUAL or NOT SURVIVABLE verdict against a proposed burn rate. Uniswap converts a WETH endowment's yield into the USDC bills are denominated in; fork tested, 10 WETH accrued 0.08638 WETH over 180 days and paid 218.698043 USDC, within two cents of oracle. Privy is how an owner signs in and commits principal, as a self-service Earn vault flow — deliberately not the only way in, because we fall back to any injected wallet, and gating the whole write path on one vendor's key turns a missing env var into a dead demo. Testing: 23 tests. 10 offline unit and fuzz tests run in 477ms with no RPC; 13 fork tests run against live Aave v3 on Base at a pinned block, no mocks. The fork tests found four real bugs that reading the code did not, including the rounding leak above. Built with Claude Code as the implementing agent. Full disclosure of what was AI-written, what the human contributed, and the eight bugs the AI introduced and later caught is in AI-USE.md in the repo.