tokentab turns your local AI coding logs into a cost report

tokentab reads the session logs three coding harnesses already write to disk and turns them into a cost statement. It parses ~/.claude/projects/**/*.jsonl, ~/.codex/sessions/**/rollout-*.jsonl and ~/.gemini/tmp/**/session-*.json, with partial support for Cursor. No API calls, no new credentials. The tool was created September 7 and sat at 481 stars as of September 12.

The token counts come straight from the logs, not an estimate. Pricing is a hand-maintained table in tokentab/pricing/prices.py, dollars per million tokens per model. Model names get fuzzy-matched, so a dated string like claude-opus-4-6-20260514 resolves to the base model claude-opus-4-6 without the maintainer having to hardcode every release tag. Cached tokens get deducted from the total so they don't show up twice in the bill, which matters if you're running long sessions where most of the context is cache reads.

Output comes in three shapes: a terminal table for a quick check, JSON for scripting, or a monthly statement at localhost:4747 if you run it with -web.

If you're paying for Claude Code, Codex, and Gemini sessions across multiple projects, there's no single vendor view that adds those up. tokentab is a workaround built from the fact that all three tools already log everything needed to reconstruct that number locally. It doesn't cover Cursor fully yet, and a hand-maintained pricing table means it lags whenever a vendor changes rates until someone updates the file. A stale price in that table is a silently wrong invoice, and the repo doesn't say how often the maintainer refreshes it.