Features / Budgets and cost control
LLM cost control that stops overspending before it happens.
AI agents can spend money faster than anyone can watch a dashboard. ACPpotamia puts a monthly budget in front of every call to OpenAI, Anthropic or your own model. Each call's cost is reserved before the call is sent, so many agents running at once can never spend more than the budget together.
How it works.
Reserved before it is sent
Input cost plus the output the call may produce is held from the budget first, then replaced by the real cost the provider reports.
Shortened, not surprised
When the remaining budget cannot pay for a full answer, the answer is shortened to what is left. When nothing is left, calls stop with HTTP 402.
Real list prices
Costs use each model's published price per million tokens, including cached input, for Claude, GPT, o-series and embedding models.
Token limits too
Set a monthly token limit alongside the dollar budget for teams that budget in tokens.
Per key, customer and task
Give an API key, each of your customers, or each task its own budget, all checked together before the call is sent.
Questions.
Can ACPpotamia cap OpenAI spend per month?
Yes. Set a monthly budget in dollars; calls that would go over it are shortened or refused before they reach OpenAI.
Does it work for Claude API costs too?
Yes. Anthropic calls are budgeted the same way, including prompt-cache reads and writes.
What happens when the budget runs out?
Calls get HTTP 402 with a clear message until you raise the budget or the month resets on the 1st (UTC).
Related features.
Rate limiting
Rate limit every AI agent's calls to OpenAI and Anthropic per team, at the edge. Requests over the limit get HTTP 429 with Retry-After before they cost money.
Loop detection
Detect and stop AI agents stuck in loops: repeated tool calls, retries and identical prompts. ACPpotamia stops the agent in seconds, before it burns your budget.
Audit logs
Every AI agent call logged: model, tokens, cost, outcome and timing, blocked calls included. Searchable for 365 days, with optional request and response logs.