Product

Repayd

On-chain deposit insurance for AI agent wallets on Arc. A policy wallet classifies every agent instruction as ROUTINE, ELEVATED, or VIOLATION before it settles — suspicious USDC is held for a two-minute containment window instead of leaving the account, clear violations are blocked before they ever reach the mempool, and whatever still slips through is reimbursed from a mutual USDC pool in the same block as the verdict. A rolling keccak hash-chain in the agent SDK records the origin of every instruction, so an owner cannot drain their own wallet and file a covered attack.

Repayd

Overview

On-chain deposit insurance for AI agent wallets on Arc (Circle testnet). Everyone is giving agents wallets; nobody is insuring them. Repayd is the safety layer between an agent's spending key and the money it can move — suspicious transfers get contained instead of settled, and money that is lost anyway comes back.

The design law: the AI narrates, the code decides. No model sits in any decision path — every tier, verdict, and payout is deterministic arithmetic anyone can re-run and compare digests against.

How it works

  • Three lanes in the GuardAccount. ROUTINE executes instantly, ELEVATED is held for a two-minute containment window, and VIOLATION is blocked pre-broadcast — the transaction is never constructed, so it never reaches the mempool and a hold never fails open.
  • Sealed, verifiable verdicts. A watcher engine signs an EIP-712 verdict over the on-chain policy hash. Every verdict is ECDSA-verified against the watcher key, malleability-guarded, freshness-windowed, and pinned to a policy version, so a verdict cannot be replayed against a changed rule.
  • Same-block recovery. A covered claim is settled from the MutualPool's junior and senior USDC tranches in the same block as the verdict — a live Arc testnet run paid out $135 against the loss that triggered it.
  • Priced from behaviour, not guesswork. Coverage is quoted by telematics: the Graph risk subgraph (38 events, 12 entities) is turned into streak, attempt, claim, and anomaly multipliers with provenance labels attached, so a safe agent's premium falls every week it behaves. One live run priced a 2.7× multiplier at $135/month off indexed events alone — no mocked data anywhere in the consumer chain.

The fraud moat

The deepest hole in agent insurance is the owner attacking themselves: drain the wallet, report a hijack, collect the payout.

Every instruction the agent receives is folded into a rolling keccak hash-chain — keccak(prev ‖ origin ‖ ownerSigned ‖ ts ‖ keccak(instruction)) — and the alibi check runs at claim time. Owner-signed → DENIED: ordering the attack is the confession. External → COVERED. computeEntryDigest is exported so anyone can re-run the check against public data, and tampering breaks the chain visibly.

Invariants and live proof

Not a mockup. The protocol ran live on Arc testnet across 14 transactions, every hash independently re-fetched and verified.

  • 102 Foundry tests across 11 suites, including 1000-run fuzz invariants: fund conservation, custody never locked, pool solvency, and a nullifier per breached transaction hash so no claim can be filed twice.
  • ERC-8004 identity 894341 registered on the canonical identity, validation, and reputation registries, with the verdict mirrored as validationRequest → validationResponse — score 25 (COVERED), responseHash matching the on-chain VerdictAccepted digest.
  • Attacks survived live. A $150 transfer to a three-day-old wallet was held and never settled (attacker balance $0). A $900 admin-override attempt was blocked pre-broadcast — no transaction exists at all.
  • Machine-to-machine payment rails. An x402 paywall over Hedera testnet settled two real paid requests through the Blocky402 facilitator, with HCS consensus-service audit receipts and Scheduled Transactions carrying recurring premiums.
  • A résumé agents can actually read, published as atlasrepayd.eth on ENSv2 with the full protocol record in Sepolia text records.

Tech stack