Deploy an AI agent that trades tokenized US equities around the clock — with a spend cap, a whitelist and a kill switch written into a contract, an operator key you made, and a track record anyone can recompute from chain events.
CAsoon on ponsThe runner reads price-feed momentum, on-chain flow and social mentions, asks a model for a structured decision, and passes it through two layers of rules before a key ever signs.
Momentum comes from the Chainlink feed UsdValuator charges against. Flow is what other agents actually filled. Social is context, never a trigger.
Claude returns schema-validated actions with a confidence and a reason — or a transparent momentum rule does, when no key is set. It never sees a wallet.
Stop-loss, take-profit, size and cooldown in the runner. Cap, whitelist, slippage and kill switch on chain, where a compromised runner cannot reach them.
“Deploy in five minutes” is the easy part. The hard part is being able to say exactly what the agent cannot do, and where that is enforced. Every claim on this page names the contract or the process that makes it true.
This is the same check your agent hits on every trade.
The router would proceed and the spend counter updates onchain in the same transaction.
Limits live in the contract, not in a dashboard someone has to be watching. A rejected trade does not revert the transaction — it completes with a logged reason, so the attempt still counts against the agent's record instead of disappearing.
Thirteen contracts, one path, same order every time.
Spend cap, whitelist, slippage ceiling and kill switch — checked before a single token moves.
Sends the approved order to the DEX, takes 0.1%, and writes AgentTrade or TradeRejected either way.
Clean fills raise the score; violations cut it. The score is recorded on chain; the suggested cap multiplier is advisory — PolicyVault does not apply it, the owner does.
Opens and closes AMM positions under the same guardrails, storing raw entry and exit amounts for honest PnL.
The staker share of every fee lands here as USDG and credits all stakers in one write. Each staker chooses the stock they are paid in; rewards are swapped at claim.
A slice of every fee builds a mutual reserve for verified protocol faults, capped per incident.
Live on chain: policy-checked swaps, onchain reputation, fee-funded staking and an MCP server exposing 32 tools (see the API). Also live: the LP routers for v2 pools and pons v3 pools. Built and tested, not deployed: the Uniswap v4 router. Not integrated: meta-transaction relaying and unattended smart-wallet execution. None of the contracts have been externally audited.
Agents trade tokenized stock, so the fee arrives as tokenized stock. It is converted to USDG and accrues to stakers, who choose the stock they are paid in; the swap happens at claim.
Split onchain in the same transaction as the fill, by FeeSplitter — one contract used by both the swap router and the liquidity router, so no route can quietly skip the staker share. The percentages live in the contract and are read from it, not printed here. Until anyone has staked, the staker share falls through to the operator side (the treasuries) rather than failing the trade.
Point Hermes, Claude or your own loop at the endpoint and 32 tools appear. API reference.
The full list of 32 tools mirrors the HTTP routes — see the API reference.
Endpoints that change state return unsigned calldata. They never broadcast.
A server key that could unstake your funds or hit the kill switch would defeat the entire design, so no route holds one. SafetyPool claim approval is not exposed over HTTP at all — that runs from the multisig.