AI trading agents

Autonomous strategies running on-chain. Each agent owns a BasketVault, picks tokens from a candidate universe, sizes weights, and rebalances on a cron. You deposit into the agent's vault and let the strategy manage your slice.

Protocol agents

The three flagship agents — same vault contract as manual vaults, but with autonomous keeper logic:

Each runs: candidate pick → rug-screen → basket compose → multi-DEX cascade swap → trade ledger.

Rug-screen

Every candidate token passes through a deterministic on-chain audit before the agent commits capital. Hard blocks auto-disqualify the token (the agent never buys it):

• Mintable supply — token bytecode contains an owner-callable mint() or _mint(). Owner can dilute holders post-launch.
• Pausable transfers — token uses OpenZeppelin's Pausable / whenNotPaused modifier. Owner can freeze the token, locking the vault out.
• Blacklistable transfers — token has a blacklist / denylist / isBlocked mapping. Owner can blacklist the vault address after the buy.
• Serial rugger — deployer wallet has ≥3 prior rugs on Base per Etherscan history.
• Thin liquidity — deepest WETH-paired pool across V3, V4, Aerodrome V2 and Slipstream has < 2 WETH (~$5k). Catches ghost pools.
• V4-only routing (toggleable) — env-gated rejection of tokens whose only DEX path is V4 (set RUG_ACCEPT_V4_ONLY=true to allow).

Soft signals are surfaced as metadata but do NOT auto-reject — a meme with concentrated holders can still moon:

• Top-holder concentration (% of supply held by single wallet)
• LP lock status (unlocked LP = exit risk if dev pulls)
• Deployer wallet age (< 7d new wallet is a yellow flag)
• Contract verification status (unverified ≠ scam but worth flagging)

The hard-block layer is pure on-chain + Etherscan-deterministic. Claude is layered on top only for soft signal context synthesis (= "does this deployer look credible / is this a real project") — never for the binary block decision.

What the rug-screen does NOT catch

Being honest about gaps:

• Honeypot via 99 % sell tax. A token with a transfer fee that effectively prevents selling is not pausable/blacklistable per se — slips through. Mitigation: liquidity threshold + volume-24h filter already eliminate most of these because they can't sustain volume. Planned: a simulateSell eth_call pre-pick (= "buy $1, sim sell, expect revert" detection).
• Brand new deployer, first rug. Deployer history check needs ≥3 prior rugs to fire — a wallet's debut rug passes.
• Post-buy rugpull. Dev removes liquidity AFTER the agent's pick lands. Catch via the vault-level redeem path (V12+ soft-skips non-standard transfers so the user can still withdraw everything else).
• Proxy upgrade-then-mint. Ownership renounced but contract is a proxy whose implementation can be swapped. Not detected.

Custom agents

Deploy your own autonomous strategy with configurable filters, signal weights, exit ladders, cadence, and fees. Full breakdown: Agent launchpad.

Depositing into an agent

1. Open /agents/<slug> (e.g. /agents/narrative)
2. Click BUY, sign one tx
3. Get ERC-4626 shares; agent strategy manages your slice
4. Withdraw anytime via the agent's vault page

Your share value moves with the agent's basket performance — minus fees on entry + exit.

Risk

• Not financial advice. Memecoins can rug. Strategies can lose money. Oracles can be misconfigured (temporary NAV display issues, vault stays redeemable).
• Sanity bound defense. V8 vaults cap each token's NAV contribution at 10k ETH to defend against oracle overflow attacks. Vault stays usable even with one broken-oracle token.
• Withdraw anytime. Shares are ERC-4626 — burn them, get your ETH back. No lockup.

Start small. Watch the agent's recent PnL on its public page. Withdraw at any time.