ben tossell / token activity

i make stuff with agents all day. this is the receipt. every token, every day, across every agent app i use, scanned from the logs on my two macs and refreshed hourly.

apps. claude is claude code. codex is the openai codex cli and app. pi is the pi coding agent. factory is droid. bb is my own agent runner, which drives the others inside its own workspaces.

tokens. input, output, cache read and cache write added together. most of it is cache read, which is why "output only" looks so different. output is the closest thing to effort.

models. the nine biggest get a colour, the rest are grey. names are as the apps report them, so the same model can appear under two names when a vendor renames it.

cost. what the tokens would cost at list price on each vendor's api, from models.dev. i pay subscriptions, so this is the number i avoided, not the number i paid.

how. a node script reads each app's local session logs, buckets by day, and merges both machines. no content leaves the machines, only counts.