Letting software spend money without a person approving each purchase is a design problem, not just a billing one. For generative media, where one video can cost a meaningful amount and an agent can request hundreds, the question is: what is the worst case, and who decides it?
Why not a subscription or a stored card
A subscription charges on a schedule regardless of use. For an agent that runs occasionally it wastes money; for one that runs a lot, a flat plan usually comes with quotas and throttling, so you end up managing limits anyway.
Usage billed to a stored card has no ceiling unless one is built. A bug, a loop or a leaked credential turns into a bill, and the card is attached to the account that owns it.
A prepaid wallet inverts this. The agent can spend only what has already been added. The worst case is a number you chose when you topped up.
What TurboMax adds on top
Prepaid credits are the foundation; the server layers further limits so the worst case is usually far smaller than the balance.
- Scope. A read-only key cannot spend at all. A generate key can, and creating one requires a paid account.
- Per-call ceiling.
generate_mediaacceptsmaxCredits. The server compares it with the exact amount it would hold and refuses first, charging nothing. - Daily caps. The account's daily credit and generation limits apply to keys exactly as they apply to a person using the website.
- Rate limits. Generation, uploads, chat and polling each have their own per-minute limits.
- Atomic debit. Credits are reserved in one step before a job is dispatched. Two concurrent requests cannot both succeed on funds that only cover one.
- Fast revocation. Keys expire (30 days to one year) and can be revoked from the app immediately.
What a key can never do
A key reaches only a short list of routes: browse models, read the wallet and history, poll jobs, upload sources, generate, cancel, retry and chat. It cannot reach account settings, billing, administration, public publishing or history deletion — even if the owner is an administrator. A leaked key therefore cannot change the password, change the payment method, close the account or expose someone else's data.
Honest accounting
Money-handling systems should be checkable. On TurboMax every hold, refund and settlement appears in the wallet ledger, each generation records its cost, and a failed, timed-out or cancelled job is refunded automatically. The price an agent is quoted is the price that is held; if the provider's real cost is lower, the difference comes back when the job settles.
A sensible setup
- Start with a read-only key while an agent is being developed.
- Move to a generate key only when needed, and give each agent or environment its own key so one can be revoked alone.
- Keep the wallet balance to what you would accept losing, and top up deliberately.
- Require
maxCreditsin the agent's instructions for every generation.
Setup steps are in connecting Claude, Cursor and VS Code; the agent workflow is in how an AI agent generates video through MCP; and the general pricing model is explained in AI video generation without a subscription.