Why I built it
I run AI coding tools all day: Copilot CLI in a few terminals, agents doing background work, Claude Code sessions on top. Each one burns tokens, and none of them shows you the meter while it runs. Finding out what a day actually cost meant a dashboard somewhere, a login, and a delay.
The thing is, the data is already sitting on the machine. These tools write session logs and transcripts locally, with exact token counts in them. Nobody was reading them.
So TokenTray reads them. A number in the tray, a popup when you want the breakdown, a local history so you can see the week. It never makes a network call; it parses log files on your own disk and keeps its history in a local database. The transcripts it reads contain your conversations, so the parser is deliberately content-blind: it extracts token counts, timestamps, and model names, and never touches message content.
I keep it running on every machine I work on. When a 1.3-billion-token day shows up on the chart, I want to know the same morning, not on an invoice.