Track your AI token usage by day and by model
Claude Code and OpenAI Codex both keep local usage logs with the model and token counts for every response, so you can chart your AI usage without any API access. Caldera's prompts have your AI agent total those logs into tokens per day, per model and per provider (counts only, never prompts or costs), and the AI usage section of your profile shows the daily chart and your model mix.
Updated 2026-10-11 · by mjonesone, LLC
Where the numbers come from
- Claude Code:
~/.claude/projects/**/*.jsonl- assistant messages includemessage.modelandmessage.usage(input, output, cache read and cache write tokens). Count each message id once. - Codex:
~/.codex/sessions/**/*.jsonl-token_countevents carry per-turn usage; the model comes fromturn_context.
What surprises most people
Agent sessions re-read cached context constantly, so totals are dominated by cache reads. On one heavy user's profile, about 7.7 billion tokens over 45 days were almost all cached context. Charting it by model also shows how much work each model actually does.
Questions
Does this show my costs?
No. Only token counts and model names are published; money amounts are deliberately excluded.
Does it read my prompts?
No. Only usage fields (model, token counts, timestamps) are totaled.
Which tools are supported?
Claude Code and Codex out of the box; any tool that keeps local usage logs can be added by your agent.