Show HN: I built a tool to diagnose why my Claude quota vanished in 10 minutes

Token auditing, usage forensics, and limit diagnosis—all powered by the session logs already residing on your local machine. Privacy first: nothing ever leaves your device.

📦 What is `tare`?

In weighing, tare is the weight of an empty container, which is subtracted to determine the weight of the actual contents.

Similarly, most of the cost in a Claude session is the "container"—the context that is re-sent repeatedly. tare subtracts that overhead to show you exactly what is consuming your tokens.

How it Works

tare enables Claude Code to analyze its own internal logs, allowing you to simply ask the agent where your usage went.

🚀 Getting Started

Installation

You can get up and running with a single command in your terminal:

npx skills add kelviq/tare -g -y --copy --agent claude-code

Setup Checklist:

Run the installation command.
Start a new Claude Code session.
Type / to verify that tare appears in the list.

[!NOTE] If tare doesn't appear, it may be because Claude Code has never installed a skill before. Check INSTALL.md for manual installation, team-wide deployment, or plugin options.

🔍 How to Use It

You don't need to memorize specific syntax; you can talk to Claude Code like a human. If you're frustrated about your limits, just complain!

Natural Language Examples

When hitting a limit: "Why was I locked out ten minutes into my session?" or "Why did I hit my limit yesterday?"
Tracking tokens: "Where did my tokens actually go this week?"
Planning ahead: "Is it safe to start a massive refactor now, or is my 5-hour window too full?"
Auditing: "Is there a background process running Claude Code?" or "Did using /clear actually reduce my costs?"
Financials: "What would this cost if I paid for the API directly?"
Reporting: "Create a redacted summary I can post publicly."

Command-Line Shortcuts

While plain English works, /tare provides specific variants for common tasks:

Command	Purpose
`/tare full diagnosis`	Comprehensive analysis of token destination and cause.
`/tare usage`	An attribution-focused version of the standard `/usage` panel.
`/tare window`	Checks the current state of the 5-hour rolling window.
`/tare report [days]`	Generates and opens an HTML usage report in your browser.
`/tare tools [days]`	Identifies which tools are bloating your context.
`/tare week`	Compares current weekly usage against the previous week.
`/tare share [days]`	Creates a redacted, privacy-safe summary for public sharing.
`/tare [question]`	Pass any specific question as an argument.

🧠 The Logic: Beyond "Naive Counting"

tare doesn't just provide a wall of numbers; it provides a cause.

~~Naive counting~~ of logs often inflates totals—by as much as $\approx 86\%$ in some datasets—because Claude Code's log format repeats API responses. tare corrects this.

Key Insights `tare` Uncovers:

The "Early Read" Penalty: A file read at the start of a long session is re-sent with every subsequent message. tare charges the tool for what it caused, not just what it returned.
The Rolling Window: Limits don't reset the moment you stop working. Work from four hours ago still counts toward your current quota.
Human vs. Script: If you see hundreds of short, parallel sessions, tare will flag this as a script rather than a human user.

Example Diagnosis:

"99% of yesterday's usage came from a tool you're running, not from you. Something spawned 1,553 short Claude Code sessions in your website project—9,022 requests, with up to 51 sessions running simultaneously. Each fresh session rebuilds context from scratch, which is the most expensive way to spend tokens..."

🛡️ Privacy & Security

Everything runs locally. No data is uploaded to external servers.

When using /tare share, the tool generates a summary containing only:

Totals
Dates
Tool names

It explicitly excludes prompts, file paths, file contents, commands, and account identifiers. You can verify these claims yourself by checking SECURITY.md and running a grep command on the source.

tare screenshot

🛠️ Technical Requirements & Dev Info

Requirements

OS: macOS or Linux.
Runtime: Python 3.9+ (pre-installed on macOS; no extra packages required).
Scope: Currently supports Claude Code logs only.

For Developers

If you prefer to work outside the Claude Code agent, you can use the three dependency-free Python scripts found in the repo. Refer to CLI.md for:

Custom scripting and cron jobs.
CSV exports.
Live per-request telemetry.

Future Roadmap:

Implementing a live TUI.
Adding Windows path support.
Aggregating anonymized summaries to identify global usage patterns.
Improving token estimation for tool results.