claude-watcher

Understand your Claude Code costs and usage patterns. Your data never leaves your machine — everything reads from ~/.claude/ locally.

Run

npx claude-watcher

That's it. No install step. npx downloads and runs the tool in one command.

Requires Python 3 (any version ≥ 3.8) — the tool installs streamlit and pandas automatically if they're missing.

Opens a dashboard at http://localhost:7755.

To use a different port:

PORT=8080 npx claude-watcher

Or install globally

If you run it often:

npm install -g claude-watcher
claude-watcher

What it shows

• How you use Claude — which tools Claude invokes most, what you ask about, when you're most active
• Where the money goes — cost breakdown by token type and model
• Key cost drivers — automatically flagged patterns that are inflating your bill
• Break & continue — sessions where you returned after a break and carried old context forward
• Idle resets — turns where the cache expired and had to be rewritten at full price
• Large early content — files or logs pasted early that rode in context for the whole session
• By project / day — spend across your different codebases

Privacy

This tool reads only the JSONL files Claude Code writes to ~/.claude/ on your local machine. No data is sent anywhere. No API calls are made. No telemetry (Streamlit's usage stats are disabled).

How costs are calculated

Costs are API-equivalent estimates based on token counts logged by Claude Code and Anthropic's published pricing. If you're on a subscription plan (Pro/Max), you are not billed per token — but these numbers show what your usage would cost at API rates, and where the heaviest usage is concentrated.