Brain Budget
An autonomous research agent that buys its own Venice AI inference, one signed x402 micropayment at a time, under a scoped ERC‑7710 delegation you can revoke, relayed gaslessly by 1Shot, live on Base.
Videos
Tech Stack
Description
What is BrainBudget?
BrainBudget is an autonomous AI research agent that has no API keys and no custody of your money. You answer a question by granting the agent a small, scoped, revocable on‑chain budget, a MetaMask ERC‑7710 delegation signed from your smart account. The agent then researches your question by paying per request for its own Venice AI inference over x402, where every single LLM call is an on‑chain‑enforceable USDC micropayment authorized by your delegation, relayed gaslessly through the 1Shot permissionless relayer, with a live spend‑vs‑budget meter and receipts you can verify on BaseScan.
The problem it solves
AI agents need spending power to be useful. But the two things people do today are both broken:
Today's "agent wallet" | What goes wrong |
|---|---|
Give the agent a private key / seed phrase. | That's custody — the agent (or whoever compromises it) can drain everything. |
Give the agent a credit card / API key. | Same problem one layer up: unbounded, hard to revoke, leaks into logs. |
Approve every action manually. | Then it isn't autonomous — you're the bottleneck. |
BrainBudget's answer is the third option that didn't exist before smart‑account delegation: scoped delegation. The user stays fully self‑custodial. The agent receives a delegation — a signed, on‑chain‑enforceable permission with caveats: spend at most $X USDC, in at most N calls, expiring in 24h, only through this payment route. The MetaMask Delegation Framework's caveat enforcers reject anything outside that scope at the contract level — not by trusting the agent's code. The delegation fails closed.
Use cases
- Pay‑per‑use AI agents that can't be over‑charged: budget is a cap, not a credential.
- Agent marketplaces / tool calling where every paid tool is an x402 resource and the buyer's authority is a revocable delegation.
- Multi‑agent teams (orchestrator → specialists) where each sub‑agent gets a strictly smaller slice of the parent's budget via redelegation.
- Gasless onboarding of brand‑new wallets that hold only USDC (no ETH) — the EIP‑7702 upgrade itself is relayed and paid in stablecoin.
- Privacy‑preserving research — inference runs on Venice's no‑data‑retention layer, paid anonymously by wallet signature rather than an account/API key.
How to use it effectively
1. Fund a user wallet with a little USDC (the demo uses ~$10) and a gateway wallet with ~$5 (one‑time Venice top‑up). No ETH anywhere.
2. Set the budget slider to the maximum you're willing to let the agent spend on a single question. That number is the delegation cap — the agent can never exceed it, even if its code is buggy or hostile.
3. Ask a research question and watch the cognition ledger: each thought is a paid call, each payment is a BaseScan link, the meter ticks toward your cap, and when the cap is reached the agent stops gracefully (or, if it tries to overspend, the caveat enforcer rejects it on‑chain).
4. For multi‑agent depth, leave the critic configured: you'll see the orchestrator hand a narrowed sub‑budget to a second agent that pays for its own review.