> ## Documentation Index
> Fetch the complete documentation index at: https://www.cloudtag.tech/llms.txt
> Use this file to discover all available pages before exploring further.

# Set a spend limit

> All spend is billed to the organization's Anthropic API key. Set the monthly limit for the organization, the default per group, and a group's own limit, and see what happens when one is reached.

Every agent in an organization runs on the organization's own Anthropic API key, so all spend is billed to that Anthropic account. Cloud Tag adds per-month limits on top, the same three tiers as Claude Tag.

| Limit | Set by | Where |
| --- | --- | --- |
| **Monthly limit for the organization** | Owners and Admins | Organization page → **Spend limits** |
| **Default monthly limit per group** | Owners and Admins | Organization page → **Spend limits** |
| **A group's own limit**, which replaces the default | Group admins, up to the default; Owners and Admins, any amount | Group's **Usage & activity** page |

- A group admin can set the group's limit up to the organization's default per-group limit. Only the organization's Owners and Admins can set it higher.
- A group's limit left empty means the organization's default. Leave the organization's fields empty for no limit.
- Months run from the 1st, in UTC.

## What counts

Each reply's cost is recorded when it finishes: the session's reported cost, or an estimate from its token counts at list prices. When an [auto-approve rule](https://www.cloudtag.tech/docs/admins/add-connections.md#auto-approve-rules) is checked, the checker's model call counts toward the reply too. When the session reports no cost, the estimate leaves out web search and other non-token charges. The group's **Usage & activity** page shows this month's spend against its limit, and the cost of each recent turn. The organization page shows the organization's total.

The final turn state, charge and queued reply are recorded together. If final usage or accounting is unavailable, the turn stays open and **Recent turns** shows a waiting message; accounting is retried without resubmitting the agent request. Interrupted runs wait for the session to stop before recording usage, including costs incurred before a failure, timeout or revoked DM access. Spend totals include those charges once recorded.

## Warnings

At 75%, 90% and 100% of a limit, Cloud Tag emails:

- for a group's limit: the group's admins and the organization's Owners and Admins;
- for the organization's limit: the organization's Owners and Admins.

Each step is sent once a month.

## When a limit is reached

- **A group's limit:** that group's agent stops answering. In each thread where someone asks, it says the group has reached its monthly budget and where an admin can raise it (at most once a day per thread).
- **The organization's limit:** every agent in the organization stops answering, with the same kind of note.

Replies resume when the month resets or someone raises the limit. Nothing is queued in the meantime: ask again once it's raised.

## Running replies

Spend is recorded as each reply finishes, so several replies running at once could pass a limit together. Under a group's limit, Cloud Tag runs at most 3 of its replies at once, and under the organization's limit at most 20 across the organization; more wait their turn and start as others finish. A reply waiting for an approval doesn't count, so several approved at once can briefly run past that number. Deployments can change these numbers with `MAX_RUNNING_TURNS_PER_GROUP` and `MAX_RUNNING_TURNS_PER_ORG`.

A running reply is checked against the limit about once a minute, counting what it has spent so far, and again before Cloud Tag answers one of its own tools for it or delivers an approval decision. One that takes its group or organization past the limit is stopped, its cost so far is recorded, and the thread gets the budget note. So a limit can be passed by at most what the replies already running spend before their next check.

## Fast mode

`!fast` threads cost more per token. When the session does not report a cost, the token-count estimate includes the fast-mode premium for input, output and cached tokens. Only an Owner can allow fast mode for the organization. See [Choose the model](https://www.cloudtag.tech/docs/users/models.md#fast-mode).

## Watching costs

Run a pilot with a limit set, then look at each group's **Usage & activity** page to see which threads cost the most. You can also see sessions and spend in the Anthropic Console for the key's organization.

Deleting a group does not undo its incurred spend: it still counts toward the organization's limit for the current month.