When a Claude agent will not stop spending

Organization accounts only

Anthropic’s Admin API is unavailable on individual accounts. You need an organization, an Admin API key (it looks like sk-ant-admin…), and permission to read usage. If you paste a normal inference key, the usage report will refuse you. BillCurfew surfaces that instead of pretending the sync worked.

What the usage report returns

GET /v1/organizations/usage_report/messages accepts bucket widths of 1m, 1h, and 1d. Group by api_key_id and model. Each result carries uncached input tokens, cache reads, 5-minute and 1-hour cache writes, and output tokens. It does not carry a dollar amount. BillCurfew prices those tokens from Anthropic’s published list, including cache writes, which are the part people forget when a long prompt is cached and then rewritten all night.

Turning the key off

POST /v1/organizations/api_keys/{id} with { "status": "inactive" } disables the key. The same call with status active turns it back on. That is the right kill switch for a contractor’s agent or a staging loop: reversible, scoped to one key, and independent of the workspace.

Where the money actually goes

• A tool-using agent that retries on a 529 can multiply output tokens without a new prompt you would notice.
• Cache writes on a 1-hour breakpoint cost more than a cache read. A job that rebuilds the prefix every call pays the write price on repeat.
• Batch and flex tiers are cheaper. If you are not on them, do not assume the usage chart’s token count maps to yesterday’s rate.

A daily cap on the contract-reviewer key, plus a spike rule, is enough to page you while the loop is still a story you can tell, not a line item you have to explain.