Transcriptmeter — Development skill for Claude Code

transcriptmeter — what your agent sessions cost, from the transcripts on your disk

Claude Code writes every session to `~/.claude/projects//.jsonl` with the API's `usage` on every response; Codex CLI writes `~/.codex/sessions/…/rollout-*.jsonl` with a `token_usage_record` per response. Everything you need to know about what a session cost is already there — tokens by class, how much of the prompt came from cache, how big the context got, what the tools returned. transcriptmeter reads it back as numbers. Nothing leaves the machine; it keeps **sizes and identifiers only** and never shows message text or a command beyond its first word or two.

npx transcriptmeter                  # totals across sessions, both harnesses
npx transcriptmeter sessions --since 7d
npx transcriptmeter session 
npx transcriptmeter tools --cap 8000
npx transcriptmeter prices           # the list-price table and its date

**Status: 0.1.0, measured and cross-checked on one machine.** The readers follow the field shapes of Claude Code 2.1.280 and Codex CLI 0.155.1 as read on 2026-09-23. The deduplication rule that makes the Claude Code numbers right — one API response is written as several lines with the same `requestId` — is the kind of fact this tool depends on; `CLAUDE.md` lists them all with their dates.

What it reports

Number	From	Why it matters
Tokens by class: uncached input, cache read, cache write (5 min and 1 h), output	usage on each response, counted once per requestId	the sum is what you paid for; the split says whether caching worked
Cache hit ratio — cache read over the whole prompt	same	KPI 4 of the token-efficiency skill: over 70% in an implement session, or something in the prefix is moving
Cache misses after a warm