Blog
Notes on measuring what coding agents cost a team · RSS
What Claude Code, Codex and the other agents spend, per engineer, team, model and repo, and how to know it without reading anyone's prompts.
Written by the person building Kibble. The technical claims are the ones the product depends on; where a post says how a file is laid out or how a rate works, the collector's source is the thing to check it against.
-
How to see Claude Code cost per engineer, without reading anyone's prompts
Three ways to break one Anthropic invoice into spend per engineer, team, model and repo: a local usage tool, a proxy, or a collector that reads the logs the agent already writes and sends counts only. What each one sees, and what it costs you in privacy.
-
Where Claude Code and Codex keep their usage logs, and what a token count is worth
The JSONL transcripts under
~/.claude/projectsand~/.codex/sessions, the four token fields in them, why cache reads dominate the tokens but not the cost, and which fields a collector should never send. -
Estimated vs billed: why a coding-agent cost dashboard must never add the two
Local logs give you tokens at list price; the vendor's Admin API gives you what was charged. They disagree for six good reasons. The rule that keeps a dashboard honest: label both, let billed win the day, never sum them.