Tokens And Cost

Tokens and cost

The Tokens tab is a dashboard over Claude Code's own token history on this machine: per-model totals, a today panel, daily peak, busiest hour, and charts that grow with the terminal.

Where the numbers come from

Source	Gives
~/.claude/stats-cache.json	Claude Code's own lifetime rollup
~/.claude/projects/**/*.jsonl	live session transcripts newer than that rollup
~/.clauth/token_ledger.json	clauth's durable per-day record

Claude Code prunes old transcripts and its rollup freezes at a date, so clauth keeps its own ledger of finalized days. That ledger is what lets the dashboard keep advancing once the transcripts behind it are gone. Days already pruned before the ledger existed are unrecoverable.

The figures cover every account sharing this machine's home directory, since that is what Claude Code's store covers. A clauth start --isolated session writes into its own throwaway store and is not counted.

Period lens

t cycles the lens: lifetime, today, this week (from Monday), this month (from the 1st). It re-scopes the dashboard cards and the per-model breakdown.

Older days from Claude Code's rollup carry a combined in/out total with no cache split. A period that reaches back into those days shows a floor rather than an exact figure, marked with a badge, and cost renders as $X+.

Cost

The cost figure is what your recorded usage would cost on the pay-as-you-go API. Nobody is billing you that: it is the value of what a subscription covered.

It is computed per model, never off a blended rate, and it prices the four token classes separately: input, output, cache reads, cache writes. The c toggle changes whether cache tokens count toward the token totals; cost always counts them.

Prices come from LiteLLM's public price table, fetched daily and cached at ~/.clauth/price_cache.json. clauth loads the cache first so the tab paints instantly and works offline. A model with no matching rate contributes nothing to cost, renders as a faint dash, and puts the surrounding totals on a $X+ floor.

Model grouping

A model past a million lifetime tokens gets its own row. Smaller non-Anthropic models fold into an others row. The a menu narrows the bars and the breakdown to Claude models only, or to everything else.

Status feed

The Status tab is separate from all of this: it polls https://status.claude.com/api/v2/incidents.json every five minutes for incidents, their severity, affected components, and update timeline, cached at ~/.clauth/status_cache.json. ⏎ opens an incident's timeline, and the action menu opens it in a browser.

Sessions

clauth sessions inventories every Claude Code session on this machine, newest first: the global store plus any live isolated runtime's own store. A session's id is its transcript filename, which stays stable across resumes.

clauth sessions              # table
clauth sessions --json       # stable field set, tokens and cost null
clauth sessions --tokens     # parse every transcript for tokens and cost
clauth info latest           # resume command, workspace, storage path
clauth resume latest         # pick a profile, then resume

--tokens reads every transcript in full, so it is slow on a large store and off by default.

Message previews in the listing are scrubbed before they render: API keys, GitHub and Slack tokens, JWTs, bearer headers, URL passwords, anything under a token / secret / password / api_key key, plus long high-entropy runs, all become [REDACTED]. The redaction is render-time only. The transcript files are never modified. Session ids and workspace paths are left intact.