Give an AI a budget — not your wallet. Bounded, revocable spending powered by MetaMask Smart Accounts (ERC-7710 delegation), x402 payments, and Venice AI.
ven-AI — Give an AI a budget, not your wallet.
Today, letting an AI agent pay for things means handing over a private key or
a pre-funded hot wallet it fully controls — unbounded, irrevocable, and
unsafe to compose. ven-AI replaces custody with programmable, bounded
permission built on the MetaMask Smart Accounts Kit.
The user grants a budget by signing an ERC-7710 spending-limit delegation
with their wallet (caveat erc20TransferAmount, e.g. $5) — off-chain, no gas.
From a single natural-language request, ven-AI plans the task, redelegates
narrowed sub-budgets to specialist agents (research, copywriting, image,
video, audio, translation), and each agent pays for the service it uses
through an x402 402 → pay → retry loop. Authority can only ever narrow as it
flows down (user → ven-AI → specialists), so the sum of all sub-agent spend
can never exceed the user's cap. Specialist output is generated by Venice AI
(text + image), and every step returns delegation hashes + an activity trace
as proof.
It is a general agent, not task-specific: the capability registry is
data-driven, and users can create custom agents that the planner picks up
automatically.
Tracks: Best Agent · Best A2A Coordination (redelegation) · Best x402 +
ERC-7710 · Best use of Venice AI · (1Shot relayer scaffolded).
Live: web ven-ai-app.vercel.app · agent ven-ai-agent.vercel.app. Chain: Base
Sepolia. No custom smart contracts — integrates MetaMask's deployed
DelegationManager + caveat enforcers via @metamask/delegation-toolkit.
Built from scratch during the hackathon as a pnpm monorepo (Next.js 15 web +
Hono agent + shared TS contract):
- ERC-7710 delegation core — build delegations with granular caveats
(erc20TransferAmount spending limit + allowedTargets), redelegation that only
narrows authority, and off-chain signing via @metamask/delegation-toolkit.
- Wallet-signed grant flow — the user signs the root spending-limit
delegation in MetaMask (EIP-712); the agent exposes a stable delegate address
via GET /spike/agent and redelegates from the user's signed root. Plus a $5
free-credit fallback so the flow demos without a wallet.
- Agent orchestrator (POST /spike) — plan → root grant → redelegate per
specialist → x402 payment → generate → return a typed SpikeResult (outputs +
delegation proofs + live budget).
- x402 payment loop — real 402 → attach X-PAYMENT → retry → 200 against a
local mock seller, bounded by each specialist's sub-cap.
- Venice AI integration — live text + image generation in the main flow
(deployed agent verified returning real output).
- General capability registry + custom agents — six built-in specialists
priced per use ($0.20–$1.00); users create their own on /app/agents, merged
into the selection pool.
- Full UI — dashboard (budget meter + Grant modal), 3-column chat (previews ·
conversation · generated output/history), agents registry, all on a
hand-built warm-paper design system.
- Real budget accounting — replaced mock numbers; cap/spent now derive from
SpikeResult, shared across pages via a budget store.
- Dual-mode settlement (gating) — off-chain simulation by default; an
env-gated on-chain redeemDelegations path with a balance guard and safe
fallback so the demo never breaks.
- Deployed — web + agent to Vercel (Base Sepolia, stable delegate key,
English output, Venice live).