Day boundaries

Tokdash cuts every day at your machine's local midnight. Date ranges, the "Today" and "Yesterday" presets, the daily totals, and the activity heatmap all use the host timezone, read fresh on each request, so it follows DST.

Timestamps are stored as epoch milliseconds — absolute instants, no timezone — and the local boundary is applied when they are bucketed. Nothing in the interface displays UTC.

Why a provider's own page disagrees

Provider usage pages generally bucket by UTC midnight. OpenAI's Codex usage profile does. At UTC+1 that puts their day boundary at 01:00 your time; at UTC+8, 08:00. So work you do in that band is filed under one date by Tokdash and the neighbouring date by the provider.

Concretely, at UTC+1, Tokdash's "16 Aug" spans:

2026-08-16 00:00 local  ..  2026-08-17 00:00 local
= 2026-08-15 23:00 UTC  ..  2026-08-16 23:00 UTC

The provider's "16 Aug" spans 00:00 UTC .. 24:00 UTC. Same underlying calls, different cuts. On one measured day this was a 7% difference in the daily total; the wider your offset from UTC, the wider the band that moves.

Two other reasons the numbers will not match

Coverage. Tokdash reads the session logs on this machine. A provider's page is server-side and account-wide, so it also counts work from other machines you are signed in on, and from cloud tasks that never touch local disk. This gap cannot be closed locally, and it is unbounded.

Token definitions differ per provider. Codex reports total_tokens as input_tokens + output_tokens, with reasoning tokens counted inside output. Tokdash carries reasoning as its own figure and adds it to the total, so its Codex token count runs slightly above the provider's — on the order of 0.1–0.5% for typical usage, more on reasoning-heavy days. Costs are unaffected.

Because of these, aligning only the day boundary would not make the two agree. Tokdash does not offer a UTC view for that reason: it would produce a second number that still did not reconcile, with no way to tell which difference you were looking at.

What to do

For a like-for-like check against a provider page, compare a week or a month rather than a single day. The boundary displacement is a fixed band at each end, so it washes out as the window grows, leaving only the coverage and definition differences above.