Claude builds a per-client AI usage billing meter

Henry Alfano passes AI costs through to consulting clients. After an unexpectedly large extra-usage bill, he used Claude to design and code a billing meter that distinguishes two different levels of evidence rather than treating the vendor total as a detailed usage record.

Local Claude Code session logs form the measured pool. The ccusage tool prices those sessions at list rates, while repository folders identify clients. Other command-line tools, including Codex, are excluded. This produces client-attributable local records; it does not explain every cloud interaction.

The remaining overage becomes a separate estimated cloud pool. Alfano’s Cowork document workflows leave directive files, snapshots and reports. A Python script counts executed document batches and rewritten bytes per project. It calculates allocation independently using each measure, bills their midpoint and displays both results as a bracket. Columns reconcile to the dashboard total to the penny. Another script adds AWS charges from the separate account used for each client.

Shell aliases produce the reports, and a Claude skill with a monthly scheduled task turns them into draft invoices. Alfano approves invoices before sending. He reports a five-minute month-end process, without an independent time study.

The transferable distinction is measured local receipts versus a disclosed allocation estimate. Exact reconciliation does not make cloud allocation measured token usage, and the account does not demonstrate agreement from clients or an external billing audit.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: Blog · Henry Alfano