KARTIK CHANDRA BISWAS
← Back to blog

claude-analytic: See Where Your Claude Code Tokens Actually Went

  • Node.js
  • CLI
  • Claude Code
  • AI
  • Tokens
  • Open Source

Tokens ran out faster than expected? Don’t know where they went? This tool shows you — locally, in seconds.

npx github:onsaurav/claude-analytic

Or install globally: npm install -g claude-analytic, then run claude-analytic.

What this actually does

A small CLI that reads your local Claude Code sessions and reports token usage and cache hit rate, so you can see why your quota disappeared. No network calls. No API keys. Everything stays on your machine.

Claude Code stores every session as JSONL files under ~/.claude/projects/. The tool reads those files, tallies tokens by type, computes your cache hit rate, and prints a summary. That’s it.

Two ideas worth knowing

Two terms drive everything below: the token (the unit you’re billed for) and the cache hit (the mechanic that makes long sessions cheap). Internalize these and the rest of the report reads itself.

The token

The smallest unit of text Claude reads or writes — roughly ¾ of a word, or about 4 characters in English. Every prompt you send and every reply Claude returns is measured in tokens, and that’s what you’re billed for.

“Hello, world!” → about 4 tokens · 500-word message → ~650 tokens · medium source file → 1,000–3,000 tokens.

There are four token types, and they’re priced very differently:

TypeWhat it isPrice / 1M
InputFresh text you sent to Claude — read from scratch, not pulled from the cache$3.00
OutputText Claude generated back to you — generation is the costly half of the loop$15.00 · 5× input
Cache writeInput stored into the prompt cache for reuse on future turns$3.75
Cache readInput served from the cache instead of re-processed — this is where savings live$0.30 · 10× cheaper

The cache hit

When you keep talking in the same session, Claude doesn’t re-read your earlier messages from scratch — it pulls them from a prompt cache. Each time the cache is reused, that’s a cache hit.

A higher hit rate means you’re reusing context efficiently and paying less. Cache hit rate = cache reads ÷ (cache reads + fresh input).

Where your tokens actually went

Sometimes tokens vanish well before you expected. The usual suspects, ordered by frequency:

What happenedWhy tokens drained
Big code generation or refactorOutput tokens cost 5× input — generation adds up faster than people expect
You ran /clear oftenEach clear kills the cache. The next prompts pay full price until the cache warms back up
Hopping between many projectsEach project has its own cache. None of them stay warm
Long idle gaps between turnsCache expires. You pay the warm-up again
Edited a file Claude read early onPrefix changed. Everything after the change cache-misses

Reading your hit rate

A day spent at 90% versus 30% can be a 3–5× cost difference. The number itself tells the story:

  • 85%+ — excellent
  • 60–85% — room to improve
  • < 50% — paying full price

What you’ll see

Run the command and you get two sections: your top spenders, then the totals. Sample output:

 ┌──────────────────────────┐
 │  Claude Analytic Report  │
 └──────────────────────────┘

 Detected Claude path: C:\Users\you\.claude\projects
 Total sessions:       3

 ▶ Top 5 Sessions by Token Usage
 ┌──────────────────────┬─────┬────────┬───────────┐
 │ Session              │ In  │ Out    │ Cache Hit │
 ├──────────────────────┼─────┼────────┼───────────┤
 │ d--workspace-claude… │ 153 │ 28,150 │ 99.99%    │
 │ d--workspace-c160c1… │  69 │ 21,284 │ 99.99%    │
 │ d--workspace-claude… │  50 │ 12,826 │ 99.99%    │
 └──────────────────────┴─────┴────────┴───────────┘

 ▶ Overall Summary
 ┌─────────────────────────────┬───────────┐
 │ Metric                      │ Value     │
 ├─────────────────────────────┼───────────┤
 │ Total Input Tokens          │ 272       │
 │ Total Output Tokens         │ 62,260    │
 │ Total Cache Read Tokens     │ 4,123,997 │
 │ Total Cache Write Tokens    │ 135,468   │
 │ Overall Cache Hit Rate      │ 99.99%    │
 └─────────────────────────────┴───────────┘

 💡 Tip: High cache hit rate = lower cost

Top 5 Sessions by Token Usage — your biggest spenders, ranked, with their individual hit rate.

Overall Summary — totals for every token type and your overall hit rate. Reference pricing per 1M tokens: input $3.00 · output $15.00 · cache read $0.30 · cache write $3.75.

Three habits that save tokens

  1. Stay in one session. Every /clear is a cold cache. Resume work where you left it whenever you can.
  2. Put stable context first. The cache matches a prefix — anything that changes early invalidates everything that follows. Stable instructions and unchanging files belong at the top.
  3. Don’t ping-pong between projects. Each project has its own cache. Quick context-switches in short bursts mean none of them stay warm long enough to pay off.

Privacy

Everything runs locally. No data leaves your machine. No telemetry, no API keys, no network calls.

Free and open source under MIT: github.com/onsaurav/claude-analytic

— a small tool, kept honest.

Contact