Budgets

Budgets cap managed LLM spend for a person, a team, or the whole organization, and decide what happens when the cap is hit: a warning, a soft block, or a hard block.

Create a budget

Open LLM usage and switch to the Budgets tab. Creating and deleting budgets requires the manage-budgets role.
Select Create budget. The dialog asks you to define scope, amount, period, and alert behavior.
Pick the Scope. Organization covers everyone; Team covers one team you then select; User covers one person.
Set the Budget (USD) and Period. The period is Monthly (calendar month) or Weekly (Monday to Sunday). Spend resets to zero at the start of each period.
Choose On exceed and the Alert threshold %. The action (below) fires when spend passes the budget amount; the alert notification fires earlier, at the threshold percentage (default 80).
Select Create. The new budget appears as a status card with a progress bar showing spend against the amount.

The Budgets tab with an organization-wide monthly budget — An organization budget: monthly amount, percentage used, alert threshold, and action.

What each action does

On exceed	Effect on LLM calls
Warn only	Calls continue unchanged. Threshold and exceeded alerts are still sent, once per period.
Soft block	Calls still go through, but the proxy marks each response as over budget (the `X-Harriet-Budget-Exceeded: soft` header) with a message naming the budget, so client apps can surface it.
Hard block	Calls are refused until the period resets or you raise the amount. The error states which budget was exceeded, e.g. "Your team Sales has exceeded its monthly LLM budget of $200.00."

Two behaviors are worth knowing. First, budgets count managed spend only — BYOK usage is billed by your own provider and is not counted against budgets. Second, when several budgets apply to one person, they are checked in order: their user budget, then their team budgets, then the organization budget, and the first hard block wins. Spend checks are cached briefly, so enforcement can lag new usage by a few moments.

Each budget's History button opens its alert history: which threshold and exceeded notifications were sent, and when. Delete removes a budget immediately.

Pair budgets with the model allowlist

Budgets control how much is spent; the model allowlist controls how fast it is spent. Restricting a team's devices to cheaper models stretches the same budget much further, and keeps an accidental "expensive model for a bulk job" from burning a week's allowance in an afternoon. Allowlist controls require Endpoint AI Team or higher — see Plans & pricing.

Recommended starting setup

Start permissive and tighten with data:

One Organization budget, Monthly, set to Warn only at roughly what you expect to spend. Run it for a full period and compare against usage analytics.
Then add Team budgets with Soft block for teams whose spend you want visible pressure on.
Reserve Hard block for cases where a stopped call is better than an overrun. It interrupts people mid-task, so use it deliberately.

💡

A hard block stops spending; it doesn't refill anything. If the real problem is an empty balance rather than an over-budget team, see Credits & top-ups.