How to set a spend limit on Claude Code agents

Three bills, three ceilings

A Claude Code session can be stopped by Anthropic, by the cloud provider if you use one, or by AOSplane before the task starts. Set the one that matches the bill you are worried about. None of them is a quality check. Review and QA are separate, and they are what runs before your team sees the pull request.

A cap on one local session

Pass --max-budget-usd with a dollar amount for that Claude Code process. The session cost Claude Code estimates, including the 1.1× multiplier it applies for a data-residency response, counts toward the flag. /usage shows the session figure. It is an estimate. The invoice is still Anthropic’s.

The agent-loop docs describe how the cap is enforced. The comparison happens after a response arrives, because that is when the token cost is known. The response that reaches the cap still finishes, so the total can land a little over the number you set, by up to that response plus any subagents still running. Subagent spend counts toward the same cap. Once the cap is reached, a new subagent is refused and background subagents are stopped. Treat the number as a ceiling with a small overshoot, not as a pre-authorization.

A cap for the whole API organization

If developers sign in to Claude Code with a Claude Console account, Anthropic creates a workspace named “Claude Code” the first time. You cannot create API keys in it. It exists so you can see Claude Code spend and set a workspace spend limit on it. For organizations with custom rate limits, that workspace’s traffic still counts toward the organization’s API limits. A workspace rate limit on its Limits page can keep Claude Code from crowding out other production traffic.

On Amazon Bedrock, Google Cloud, or Microsoft Foundry, Claude Code does not send those metrics back to Anthropic. The cap lives in that cloud’s budget controls, or in a gateway you run.

A cap on a Teams or Enterprise plan

On Claude for Teams and Enterprise, each seat has an allowance on a rolling five-hour window and a weekly window, shared with chat on that plan. Usage inside the allowance is not metered in dollars. To let someone continue past it, turn on usage credits and set spend limits for the organization, a group, or one member, in the admin settings. /usage-credits is how a developer opens those settings, or asks an admin, depending on their role.

The message the developer sees tells you which ceiling it was. A session or weekly limit is the seat allowance. An individual, org, or team spend limit is usage credits hitting a number you set. A gateway message is a cap on a gateway you operate. None of those is AOSplane refusing a task.

The hard cap on one task

When Claude Code runs on AOSplane, the Anthropic controls above still apply to the account you connected. AOSplane adds a check before the run. The organization limit, the project limit, and the agent limit are read together. The tightest one is the cap for that task. If the reservation would exceed what is left, the task does not start. A run that has already started is not cancelled when a later task is refused.

That is the contrast with a shared AI-credit balance. Linear’s cloud coding agent draws on the workspace’s AI credits, and admins manage spend limits on that balance. Copilot Business and Enterprise pool AI credits across seats. Rovo Dev Standard puts 2,000 credits a month at the organization, and Code Reviewer is moving to the organization’s shared Rovo allowance. One session on those products can spend credits the next session needed. A hard per-task cap refuses the task instead. The sources for those three products are on their comparison pages: Linear, Copilot, and Rovo Dev.

The cost of a task that did run is on the task’s record, with the spec, the review notes, and QA’s results. That is what the team receives. It is not a reconstruction from a pool after the fact.

A practical order

1. Set the Anthropic workspace limit, or the Teams spend limit, so the company has a ceiling the provider enforces.
2. Pass --max-budget-usd on a local session you might leave running.
3. On AOSplane, set the agent ceiling inside the project ceiling inside the organization ceiling. The task stops at the tightest.

The team version of the same loop is Claude Code for teams. Running it from a Linear issue is the other note.

Claude Code’s session flag, workspace, and Teams controls are summarized from Manage costs effectively and the agent loop, as read on 11 November 2026. Anthropic’s pages are the authority for their own limits, and they change. AOSplane’s check is what this site says the product does today.