The genuine objection to bring-your-own-key AI is not the rate. It is the absence of a ceiling.
A subscription has a worst case printed on it. An API key does not. Somewhere between "this costs a few dollars a month" and a scripted flood of requests at three in the morning is a number nobody wants to discover from a credit card statement.
So Helpie ships the ceiling.
Three limits
Set any or all of them, in dollars:
- Daily, which catches a single bad day.
- Weekly, which catches a sustained problem you did not notice on day one.
- Monthly, which is your actual budget.
Leave one at zero and it is not enforced. Set all three and the tightest one wins.
What happens at the ceiling
The check runs before the API call, not after. When a limit is reached, no request is made at all, so the cap is a real boundary rather than a report on money already gone.
What the customer sees is up to you:
- A ticket form. The assistant explains that AI help is paused and offers to pass the question to a person. This is the default, and for most shops the right answer: the customer is still helped, just by you.
- Direct contact details. A form for getting in touch, without creating a ticket.
- Nothing. The widget disappears entirely until the period resets.
The important part is that none of these are a broken widget or an error message. Reaching a spending cap should look like a deliberate choice, because it was one.
Alerts before you get there
Helpie checks spend every fifteen minutes and emails you when you cross a threshold you set, defaulting to 75% of a cap. A second email goes out if a cap is actually reached.
Each is sent once per period, so a busy week does not turn into forty identical emails. The threshold email includes what has been spent, what is left, and which conversations are consuming the most.
Alerts go to whichever addresses you nominate, falling back to the site administrator.
Where the money went
There is a running ledger, one row per API call, recording the model used, tokens in and out, the estimated cost, and which conversation it belonged to.
That surfaces as three things on the spend screen: colour-coded progress bars for each period, the caps form, and a "top consumers today" table showing the ten most expensive conversations with their call counts.
That last table is the useful one. A single conversation with forty calls against it is not a normal customer. It is either a bug worth finding or someone playing with your widget, and either way you want to see it.
Weeks and months, defined
The week starts Monday at midnight in your site’s timezone. The month starts on the first. Both use your timezone rather than a server default, so the numbers match the days you actually think in.
The cost figures are estimates
Helpie calculates cost from token counts and published per-model rates. It is close, and it is not your invoice.
Anthropic’s billing is the source of truth, and Helpie says so on the screen rather than presenting its own arithmetic as authoritative. The estimates are there to keep you inside a budget and to show you which conversations are expensive, which they do well.
The other half of cost control
Caps stop the worst case. Caching stops the routine case from costing anything much in the first place, and a repeat question from an anonymous visitor is served without an API call at all.
Between the two, the shape of the bill is: small, predictable, and with a hard number you chose written across the top of it.