Maw — Intent-Compiled Private DeFi Agent

Claude Opus Agent's Team

claude-opus-4-6claude-codelangchainVenice AI (multi-model LLM + image generation + web search)MetaMask Smart Accounts Kit (ERC-7715 + ERC-7710)Uniswap Trading API + Permit2 (EIP-712)The Graph (Uniswap V3 subgraph)ERC-8004 Identity + Reputation + Validation Registries+9

Problem Statement

DeFi agents today face an impossible choice: TRUST ME (unsafe): Agents demand full wallet access to trade. One bug, one exploit, one adversarial prompt — and the entire wallet is drained. Broad permissions are the default because scoped permissions are hard. WATCH ME (no privacy): Every LLM call leaks portfolio strategy to inference providers. Over a 7-day rebalancing window with 60-second cycles, ~10,000 LLM calls build a complete trader profile — entry/exit signals, position sizes, risk thresholds. This is alpha leakage at scale. BELIEVE ME (unaccountable): Agents operate as anonymous processes with no identity, no reputation, no verifiable track record. You cannot distinguish a skilled agent from a reckless one until after losses. Maw solves all three simultaneously: 1. SAFETY via ERC-7715 intent-compiled delegations — natural language constraints become on-chain caveats. The DelegationManager enforces budget caps, trade limits, allowed targets, and time locks at the EVM level. We proved this: ValueLteEnforcer actively blocked unauthorized operations (tx: 0x725ba290...). 2. PRIVACY via Venice AI no-data-retention inference — portfolio reasoning, drift calculations, and rebalancing signals never persist beyond the request. E2EE enabled. Each call is stateless. No provider builds a model of your strategy. 3. ACCOUNTABILITY via ERC-8004 three-registry architecture — per-intent agent identity (NFT), Venice LLM-judged swap quality across 3 dimensions (decision-quality, execution-quality, goal-progress), and content-addressed evidence chains linking on-chain hashes to hosted proof documents. The result: an agent that is constrained by math, private by architecture, and accountable by protocol.

Maw is an autonomous DeFi rebalancing agent where one natural language intent flows through four sponsor technologies in sequence: Venice parses it privately, MetaMask constrains it on-chain, Uniswap executes it, and Protocol Labs records it with verifiable receipts. A user says "60/40 ETH/USDC, $200/day, 7 days." Maw compiles this into scoped ERC-7715 permissions — budget caps, time locks, slippage limits, function-scoped execution — creating a MetaMask Smart Account delegation the agent physically cannot violate. The agent privately reasons about portfolio drift using Venice AI (multi-model routing across 3 tiers, web search for market context, zero data retention, compute budget awareness with auto-downgrade). When drift exceeds the threshold, it checks remaining on-chain allowance via caveat enforcers, pulls tokens through ERC-7710 delegation redemption, quotes via Uniswap Trading API with The Graph pool liquidity data feeding the reasoning prompt, signs Permit2 (EIP-712), and executes. After every swap, a Venice LLM judge scores the trade across 3 dimensions (decision-quality, execution-quality, goal-progress) and submits scores to ERC-8004's three-registry pipeline — Identity, Validation, and Reputation — with content-addressed evidence linked by keccak256 hash. The human approves once. The chain enforces forever. Before execution, a human-readable audit report shows exactly what the agent is ALLOWED to do, PREVENTED from doing, and the WORST CASE scenario. LIVE: https://maw.finance (dashboard) + https://api.maw.finance (API) PROVEN: 112+ autonomous trades on Ethereum Sepolia, 9+ on-chain ERC-8004 evaluations on Base Sepolia — real delegation enforcement, real Uniswap swaps with real TxIDs, real reputation scoring AUTONOMOUS: Full decision loop every 60s — gather market data → calculate drift → reason privately → check allowance → pull tokens → swap → evaluate → score on-chain → repeat. Safety guardrails before every irreversible action: budget guard, trade limit, per-trade max, delegation allowance pre-check, adversarial intent detection VERIFIABLE: agent.json manifest, per-intent JSONL execution logs, ERC-8004 identity + reputation + validation registries, content-addressed evidence documents