Token Trick
See where your Codex and Claude Code usage goes: cached and fresh tokens, estimated cost, recorded runtime, and account quota.
Public beta · your own account and local collection client
Setup and privacySee where your Codex and Claude Code usage goes: cached and fresh tokens, estimated cost, recorded runtime, and account quota.
Public beta · your own account and local collection client
Setup and privacyLoading…
{{ state.error }} Saved readings may be out of date.
No saved quota readings for this window.
{{ describe(selected) }}
Provider quota, not raw tokens. Hourly saved observations; all accounts share a timeline ending at the latest observation. Dates are UTC. Gaps over two hours are dashed; resets and increases break the line. Readings may lag FairyStack’s live account meter.
| Account | Observed interval (UTC) | Remaining | Drop | Elapsed |
|---|---|---|---|---|
| {{ p.label }} | {{ pct(p.previous.remaining) }} → {{ pct(p.remaining) }} | {{ Number(p.drop.toFixed(2)) }} pp | {{ duration(p) }} |
Wait: every tool call only polls background work. Harvest: every call only reads task output. Notify wake: a Claude reply to a task notification. Unrecorded: usage collected before request purposes were recorded. Weights are list-price ratios, not dollars: fresh input 1, cached input 0.1, Claude cache write 1.25, Claude output 5, Codex output 8. Codex has no harvest or notify-wake class.
Collect from one computer per Token Trick account. Each upload replaces that account’s previous snapshot; multiple computers are not merged.
Requires Python 3.10+ and pipx. Run this on the computer where you use Codex or Claude Code.
pipx install https://tokentrick.com/downloads/token_trick_client-0.17.0.tar.gz
Loading key controls…
Replacing or revoking the key stops collectors using the old key. After replacement, run setup again on your computer.
{{ state.status }}
{{ state.key }}
token-trick setup --no-schedule token-trick collect
Paste your key into the hidden setup prompt. The client reads the last 90 days of local logs and uploads your snapshot. Return to Usage and refresh to see it.
Repeat token-trick collect when you want fresh data. On Linux with systemd, run token-trick setup to enable daily automatic collection. macOS and Windows use manual collection; those platforms have not yet been verified end to end.
Daily and per-model token counts, cache counts, recorded runtime, quota readings, service tier, client type, and opaque session/turn/response identifiers are uploaded to your private Token Trick ledger. Recent response metadata includes timestamps, effort, compaction counts and character counts. Prompts, response text, tool contents, local paths and raw source files are not uploaded.
Your Token Trick key stays in your local configuration with restricted permissions. Claude’s existing OAuth credential is read locally to request its quota meter; that credential is never sent to Token Trick. Missing or expired quota access is reported as unavailable. Use token-trick collect --no-claude to exclude Claude usage and quota lookup.
Data is collected from the configured local log directories, not from every device or provider account. Raw-file inspection is available only for the server’s own collector; uploaded session totals and available response diagnostics remain usable in your dashboard.
Dollar values are estimates at published model prices, not a subscription bill. Missing models remain unpriced.
Daily charts show up to 30 days from the 90-day uploaded window. Response diagnostics retain at most two days and 5,000 responses; availability depends on the log format. The public client collects Codex and Claude Code; organization-wide OpenAI API usage needs a separate server collector. Automatic scheduling currently supports Linux systemd only.