Pretend budget, nothing real. This demo runs entirely in your browser: no request is sent to any AI provider or to Xpenvo, nothing is stored, and the model names and prices are examples. The rules are the same ones the real proxy uses.
Try a budget
One agent key on a daily budget of $5.00. Send requests, start a loop, and see what gets stopped before it costs anything.
Spent
$0.00
of $5.00 today
On hold
$0.00
worst case of running requests
Blocked
0
never reached the provider
Spend avoided
$0.00
worst case of blocked requests
Worst case of one request with these settings: $0.06. Lower max_tokens and the worst case drops, so more fits in the budget.
Requests
Nothing yet. Send a request to start.
What you just saw
- Before a request is sent, Xpenvo puts its worst-case cost on hold. If that would not fit in the budget, the request is refused and never reaches the provider, so it costs nothing.
- When the answer is done, only the real cost is charged and the rest of the hold is released.
- Many requests at once see each other’s holds, so a burst cannot slip past the limit.
- A stuck loop is caught by comparing a fingerprint of each request, never its text, and the key is paused.
- Big requests can wait for a person’s approval, and one kill switch stops every agent at once.