클로드 쿼터 10분 만에 소진되어 원인 분석 도구 'Tare' 제작

tare Ask Claude Code where your usage went. You hit a usage limit and don't know why. Your quota drains faster than it used to. You suspect something is eating tokens in the background. The records that answer all of this are already on your computer — Claude Code keeps a log of every request it makes. tare teaches Claude Code to read its own logs, so you can just ask. No dashboards, no commands to learn. Install once, then ask in plain English. Tare: the weight of the container, subtracted to find what's inside. Most of what a session costs is the container — context re-sent again and again — and that's exactly what these tools subtract. Install One command, in any terminal: npx skills add kelviq/tare -g -y --copy --agent claude-code That's it. Nothing else to set up — no accounts, no packages, no configuration. Start a new Claude Code session and it's live — type / and check that tare appears. If it doesn't, see the troubleshooting note — on machines where Claude Code has never installed a skill before, the installer can miss it. (Other ways to install — by hand, for a whole team, or as a Claude Code plugin — are in INSTALL.md .) Then just ask Open Claude Code and ask your question the way you'd ask a person. These all work — the words don't have to match, complaining about your limits is enough: When you hit a limit Why did I hit my usage limit yesterday? I got locked out ten minutes into my evening session — how is that possible? Did I hit the 5-hour limit or the weekly cap? Why am I burning through my quota so much faster this week? Where your tokens go Where did my tokens actually go this week? Which of my projects is eating my quota? Which model is costing me the most? What's the most expensive file Claude keeps re-reading? How much did that giant session yesterday actually cost me? Are my MCP servers adding a lot to my context? How much overhead do subagents and skills add? Before you start something big How full is my 5-hour window right now? Is it safe to start a big refactor now, or should I wait for my window to clear? Checking for things you forgot Is something running Claude Code in the background? Was Claude Code active while I was asleep? I set up an automation last month — what is it costing me? Tuning your setup I started using /clear between tasks — did it actually help? Compare this week to last. Did the latest Claude Code update change my usage? What would my usage cost if I were paying for the API directly? What one change would save me the most? Reports and sharing Make me a usage report I can open in my browser. Give me a summary I can post publicly — with nothing private in it. Export my usage to a spreadsheet. Prefer typing commands? /tare runs the full diagnosis, and takes variants for the common asks — asking in plain words always works too: /tare full diagnosis — where tokens went and why /tare usage at-a-glance panel — like /usage , with attribution /tare window how full is the 5-hour window — safe to start? /tare report [days] build the HTML report and open it /tare tools [days] what is filling my context /tare week compare this week with last /tare share [days] redacted summary safe to post publicly /tare why did I hit the limit yesterday any question works as the argument What you get back Not a wall of numbers — a cause. The answer to "why did I hit my limit yesterday?" looks like this: 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, up to 51 sessions running at once. Your own hands-on work that day was 93 requests. Each fresh session rebuilds its context from scratch, which is the most expensive way to spend tokens... ...followed by the evidence, what to check, and what to change. And when everything is actually fine, it says that: usage proportionate, no anomaly, here's what's normal for you. Real output lives in examples/ — a shareable summary produced by /tare share , and the HTML report. What it knows that a raw token count doesn't Correct totals. Claude Code's log format repeats each API response several times over; naive counting inflates totals — by 86% on the data this was built against. tare deduplicates properly. The real cost of context. 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 — which is how one big file read early can quietly dominate a week. The rolling window. Limits don't reset when you walk away; work from four hours ago still counts. tare can tell you how full your window was at the exact moment you were locked out. The shape of automation. Hundreds of short parallel sessions is a script, not a person. tare recognises the signature and says so. Private by design Everything runs on your machine and nothing leaves it. The scripts make no network connections at all. When you ask for a shareable summary, it contains totals, dates and tool names only — no prompts, no file paths or contents, no commands, no session or account identifiers — so you can post it publicly or send it to a colleague and ask "what am I missing?" Don't take that on faith: SECURITY.md states exactly what each file reads, writes and sends, and shows how to verify every claim yourself with one grep. Requirements Claude Code on macOS or Linux Python 3.9+ — already present on every Mac; no packages to install Currently reads Claude Code's logs only, not other coding agents' Demo tare-screen-recording.mp4 For developers Everything the skill does, you can also do by hand: three dependency-free Python scripts with recipes for scripting, cron, CSV export, live per-request telemetry and more — see CLI.md . Issues and PRs welcome. Useful directions: a live TUI, Windows paths, aggregating anonymised summaries across users to spot patterns no single person can see, and better token estimation for tool results. MIT licensed.