Atlas Prime Book

Atlas is agent credit infrastructure.

It turns a MetaMask ERC-7710 delegation into bounded, financeable authority for AI agents. An owner signs once through a MetaMask Smart Account. The agent never receives custody or private keys. Inside that signed permission, the agent can borrow, open margin exposure, grant hedge authority to another agent, and liquidate lender-first when the book is stressed.

The demo is a non-custodial prime book for an AI agent: borrow → margin trade → hedge → liquidate

On Base Sepolia, Atlas runs the credit side of the book: credit line, collateral, repayment route, risk checks, and permission redemption through MetaMask’s DelegationManager.

On Arbitrum Sepolia, Atlas runs the trading side: GMX margin exposure, pinned settlement, hedge execution, and keeper-triggered liquidation.

Venice provides the private reasoning layer. The agent can read book state and propose actions without leaking strategy before execution. But Venice can't bypass the bound: every action is clamped by the signed permission.

Core primitive: the lender does not underwrite the agent’s identity or reputation, instead they underwrite the repayment path.

Atlas shows how AI agents can become financeable counterparties: borrowers, traders, hedge allocators, and risk managers under one signed permission.

We use Prime brokerage as an example to introduce non-custodial agent credit infrastructure as a new category.

We built Atlas into a working non-custodial agent credit desk: one signed MetaMask ERC-7710 delegation, redeemed through MetaMask's canonical DelegationManager, becomes a bounded mandate that an AI agent uses to run an entire prime book — borrow undercollateralized, open margin exposure, grant a downstream agent its own smaller hedge line (child authority that can only shrink, never exceed the parent), and unwind lender-first before the cliff.

One cryptographic committment, one book: financing → margin → lend onward → lender-first liquidation.

We built on the following cook-off stack:

1. MetaMask Smart Accounts + the Delegation Toolkit (ERC-7710) are the authority substrate — caveats enforced on-chain by the canonical DelegationManager, so the agent's limits are legible to a counterparty, not just to the wallet.

2. Venice AI is the agent's private brain — live inference reasoning about size/direction/hold, advisory only: the on-chain bounds hold regardless of what it argues.

The book spans two real venues on two chains.

1. The credit side is live on Base Sepolia (Euler V2 EVK/EVC, self-deployed) — every draw, repay, child-line carve, and quiet over-bound revert is a real, clickable Basescan transaction.

2. The trade side runs on GMX v2's real production contracts on Arbitrum Sepolia, with orders filled by GMX's own keepers. The loan lives on Base, the trade lives on Arbitrum, and the permission travels with the agent, not the venue.

The testnet margin bridge is badged SIM BRIDGE (production is Circle CCTP) and the margin-call price path is badged MODELED — every on-chain redeem, revert, keeper unwind, and lender-first repay is REAL.

We also built Mission Control — one persistent prime book where every act updates the same book — with a live Prime Book state view, a receipt trail of on-chain proofs, and a four-act flow: borrow → margin trade → hedge → liquidate.

The real progress was turning "agent safety" into agent financeability. For example, Agent Wallet keeps an agent safe for you while putting the bound on-chain makes its limits legible to everyone else — and that legibility is the credit. A counterparty no longer has to trust the agent or hold its assets: it can read the bound, price the repayment path, and verify execution on-chain. Agent credit is the primitive; non-custodial prime brokerage is the proof case.

Bootstrapped