Token Drain Analyzer

Your Max/Pro weekly limit dies in a day on "light" work? This reads your real Claude Code session logs and shows where the tokens actually go — then tells a genuine bug apart from expected context-driven drain. Companion to the 5-question Token Checkup; this one uses your actual data.

🔒 100% in your browser. Your logs are parsed locally with JavaScript and never uploaded — there is no server. You can disconnect from the network and it still works.

1. Pick your session logs

They live in ~/.claude/projects/ (each project is a folder of .jsonl files). Select that folder, or drag in any .jsonl files.

or drag & drop .jsonl files here

2. Where your tokens went

of your raw token volume is cache_read — context re-read on every turn

UTC day	input	cache_write	cache_read	output

Heaviest sessions (which one ate the quota)

session	messages	cache_read	output

3. Bug, or your context?

4. What actually moves it

/clear between tasks. The single biggest lever. It drops the accumulated context that gets re-sent every turn. One marathon session is what melts the quota fastest.

Scope the working dir. Open Claude Code on the subdirectory you're touching, not a giant monorepo root, so less gets pulled into context.

Watch a big early file. A large file (or generated artifact) read in early gets re-read every later turn. Keep it out unless needed.

Honest caveat: the raw cache_read count is volume, not your dollar/quota cost — cached reads are weighted far cheaper per token than output. Don't read "1.2 billion" as a bill. The value here is the attribution (which sessions, and whether they match your usage), not the absolute number.

Stop the next cost surprise before it happens

cc-safe-setup is a free set of Claude Code safety hooks — one command installs guards against runaway cost, destructive ops, and secret leaks. npx cc-safe-setup

More on why Max plans hit limits fast: the #16157 cluster, explained.