Limits and checks
Every order an agent proposes is checked twice before anything is signed: once on the server, then in your tab against the ghost's real balances. Here is each check, what the server receives, and what can still go wrong.
Checks before signing
| Check | Where | If it fails |
|---|---|---|
| The ticker is on the agent's list | Server and tab | The order is dropped and noted as not on this agent's list. |
| The size is within the largest single order | The server caps it; the tab refuses anything still above | Skipped: above the per-order limit. |
| A buy has the cash for it | Tab | Skipped: not enough cash in the ghost. USDG is used first, then ETH. |
| About $2 of ETH stays for gas | Tab | ETH spending stops short of that reserve. |
| A sell does not exceed holdings | Tab | Skipped: the ghost does not hold that much. |
| At most six orders per plan | Server | Further orders are dropped. |
| The mandate has not ended and the agent is not paused | Tab | Nothing is signed automatically. |
Skipped orders appear in the conversation with their reason. Orders that pass are quoted through LI.FI with a 1% slippage limit. Buys pay with USDG when the ghost holds enough, otherwise with ETH. Sells are paid out in USDG.
Who signs
The agent's ghost, in your tab. If a token needs an approval first, the ghost sends it, waits for it, then sends the swap, and waits for each order before starting the next. With autopilot off, each of those transactions shows a confirmation sheet you can reject. With autopilot on, only the agent's own ghost skips the sheet, and only while its plan runs. Every other address keeps asking.
The worst case
The agent can only spend what its ghost holds. It cannot reach your main wallet, your other ghosts or an approval you gave from another address. If every order it places goes badly, you lose at most the ghost's balance. Keep that balance at the amount you are willing to put at risk.
What the server receives
| Data | Ghostdex server | Anthropic |
|---|---|---|
| Mandate, agent name, allowed stocks, per-order limit, end date | Yes | Yes |
| Balances of the agent's ghost (ETH, USDG, stocks) | Yes | Yes |
| Your instruction and the last few messages | Yes | Yes |
| Current prices of the allowed stocks | Added by the server | Yes |
| Your main wallet's address | Yes, to check a $GHOSTDEX balance for the larger allowance | No |
| Your IP address | Seen as with any request, and used to count the daily allowance | No |
| The ghost's address, or any key | No | No |
A preview sends only your instruction; the server plans with $500 in USDG and a $150 order limit. The agent route keeps nothing after it answers, apart from a count of today's runs per connection, held in memory. Anthropic processes the request under its own terms. The plan comes back as structured data, and the server returns it to your tab without sending anything to the chain.
What can still go wrong
- The model can be wrong. It can misread a mandate, misjudge the market or act on a price that has moved. The checks limit how much it can do, not whether it is right.
- Prices move. Prices used for a plan can be a couple of minutes old. Each order still gets its own quote and reverts if it would fill beyond its 1% limit.
- Outside US market hours, spreads widen and pools move more for the same order.
- The agent may decline. If an instruction is unclear or outside the mandate, or prices are missing, it places no orders and says why.
- A plan can fail partway. Orders run one after another. If one fails, the failure is reported and the remaining orders still run on their own.
- Scheduled checks stop when the tab closes. An agent that must react while you are away will not.
An agent is a tool that executes your mandate. It is not advice, and Ghostdex does not manage your money.