tare

tare is an open-source Claude Code skill that teaches the AI coding assistant to read its own session logs and answer plain-English questions about token usage, quota limits, and cost attribution. Built by kelviq and licensed under MIT, it runs entirely on your local machine — no network connections, no accounts, no dashboards.

What It Is

tare is a token audit and usage forensics tool for Claude Code users. It installs as a Claude Code skill and exposes a set of Python scripts that parse the session logs Claude Code already writes to your machine. Once installed, you can ask questions like "Why did I hit my usage limit yesterday?" or "Which project is eating my quota?" and get a cause-first answer backed by evidence — not just a wall of numbers.

The name is a reference to the tare weight concept: most of what a session costs is the container (context re-sent repeatedly), and tare subtracts that to show what's actually inside.

How It Works

tare installs via a single npx skills add command and requires no packages beyond Python 3.9+, which ships with every Mac. After installation, a new Claude Code session gains access to the /tare slash command and its variants:

• /tare — full diagnosis of where tokens went and why
• /tare usage — at-a-glance usage panel with attribution
• /tare window — how full the rolling 5-hour window is
• /tare report [days] — builds an HTML report and opens it
• /tare tools [days] — what is filling context
• /tare week — compares this week with last
• /tare share [days] — redacted summary safe to post publicly

Plain-English questions also work as arguments to /tare, so the words don't need to match any command syntax.

What It Knows That Raw Token Counts Don't

tare surfaces several insights that naive log parsing misses:

• Correct totals. Claude Code's log format repeats each API response multiple times; the README notes that naive counting inflates totals — by 86% on the data the tool was built against. tare deduplicates properly.
• Real context cost. A file read early in a long session gets re-sent with every later message. tare charges tools for what they caused, not just what they returned.
• Rolling window awareness. Usage limits don't reset when you walk away; tare can report how full the window was at the exact moment a lockout occurred.
• Automation detection. Hundreds of short parallel sessions is a script signature, not a person. tare recognises the pattern and flags it.

Privacy and Security Model

Everything runs locally. The scripts make no network connections. When generating a shareable summary via /tare share, the output contains only totals, dates, and tool names — no prompts, file paths, commands, or session identifiers. The project's SECURITY.md documents exactly what each file reads, writes, and sends, and provides a single grep command to verify every claim independently.

Requirements and Platform

• Claude Code on macOS or Linux (Windows paths are listed as a useful future direction)
• Python 3.9+ (pre-installed on every Mac; no additional packages required)
• Currently reads Claude Code logs only, not other coding agents'

The three core Python scripts are dependency-free and can also be run directly from the command line, used in cron jobs, or piped to CSV export — documented in CLI.md in the repository.