Budgets at every level. Down to a single agent run.
Set a monthly budget for the organization, each team, group, person and key, add rules by the hour, day or week, and cap what any one agent run may cost. Every budget on a request's path applies, and a limit refuses the request instead of sending you a bill.
Spend you decide, not spend you discover.
Monthly budgets
For the organization, teams, groups, members and keys. A soft alert or a hard cap, with alerts at 80% and 100%.
Budget rules
Up to five per key, team or account: an hour, day, week or month, on cost, tokens or requests, to warn or to block.
Gateway Pro and upA cap on every agent run
Owners set the most one agent run may cost. A run stops before its next model call, and the CLI's --max-cost can only lower it.
Strict budgets
Make a key's budget or an organization's hard cap strict, so requests running at the same time can't spend past it together.
Cheaper by default
Fallback chains that try the cheapest or fastest model first, a default or forced model per key, and a reasoning level range per key.
Your app's end users
Name your app's end user on a request and give each one a monthly budget and rate limit. Their ids are hashed before any provider sees them.
One request, every budget it touches.
The organization, the team, the person, the key, the run.
A request is checked against every budget it falls under, and the tightest one wins. You see where each budget stands in the Portal, and in the response's headers.
- Budgets nest: org, team, group, member, key
- Each answer's cost in
x-tempr-cost-usd - Alerts by email and webhook
A day's limit for a bot. A week's for a team.
Budget rules catch the runaway loop a monthly budget would only notice at month end. Write them as short lines, on keys, teams or the whole account.
- Hour, day, week or month
- Cost, tokens or requests
- Warn, or block with
402 rule_budget_exceeded
day cost 20 blockhour requests 300 blockweek requests 5000 warnmonth tokens 40M blockhour requests 300 block · resets at 15:00
See which runs cost the most, and why.
The Usage page lists this month's costliest agent runs, and each one opens as a trace of every model and tool call with what it cost.
- Costliest runs this month, for owners and admins
- Every run traced call by call
- Usage by product, model and person
- Or ask Claude Code or ChatGPT, through the Gateway's MCP server
About spend.
Does a refused request cost anything?
No. A request a budget refuses never reaches the provider, so nothing is billed for it, by Tempr or by the provider.
What happens when a budget runs out?
A hard cap refuses further requests until the period resets. If the key has a used-up model, the request moves to that model instead, so work carries on at a cost you chose.
Do budgets cover every tool?
Yes. Every request that goes through Tempr counts, whether it comes from Tempr's agents, Claude Code, Codex, Cursor or your own apps.
Which plans include budget rules?
Monthly budgets come with every Gateway plan. Budget rules come with Gateway Pro, Scale and Enterprise.
Know what AI costs before the invoice does.
Put budgets on every team, person, key and agent run, in an afternoon.