tare

tare teaches claude-code to read its own request logs, so you can ask it in plain language why you hit a usage limit, which project is eating your quota, or whether something ran sessions while you were asleep. It ships as a skill plus three dependency-free Python scripts, reads only the session logs already on the machine, and makes no network connections tare-claude-usage-audit. The name is the weight of a container subtracted to find what is inside, and the README's framing is that most of what a session costs is the container: context re-sent over and over.

How it works

The skill routes a question to the scripts, and the scripts parse Claude Code's session logs. The README names four things it does that a raw token count does not.

It deduplicates. Claude Code's log format repeats each API response several times, and on the data tare was built against, naive counting inflated totals by 86%. Anyone running a homemade jq sum over those logs should treat that as a warning about their own numbers.

It charges tools for what they caused rather than what they returned. A file read early in a long session is re-sent with every later message, so one large early read can dominate a week. This is the same cost structure prompt-caching-in-agents explains from the cache side.

It models the rolling window. Limits do not reset when you walk away, since work from four hours ago still counts, so tare reconstructs how full the five-hour window was at the moment you were locked out.

It recognises automation by shape. Hundreds of short parallel sessions come from a script, not a person. The README's example answer finds that 99% of a day's usage came from a tool that spawned 1,553 short sessions, 9,022 requests with up to 51 running at once, against 93 requests of hands-on work. Each fresh session rebuilds its context from scratch, the most expensive way to spend tokens.

Usage

Install for Claude Code globally, then start a new session and check that tare appears under /:

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

After that, questions in ordinary language work, and so does a command form:

/tare                  full diagnosis: where tokens went and why
/tare usage            at-a-glance panel, like /usage with attribution
/tare window           how full the 5-hour window is right now
/tare report [days]    build the HTML report and open it
/tare tools [days]     what is filling the context
/tare week             compare this week with last
/tare share [days]     redacted summary safe to post publicly

The share output holds totals, dates and tool names only, with no prompts, paths, file contents, commands or account identifiers. A SECURITY.md lists what each file reads, writes and sends, with a grep to verify it. CLI.md covers using the scripts directly for cron jobs, CSV export and live per-request telemetry.

Limitations

It runs on macOS and Linux only, needs Python 3.9+, and reads Claude Code's logs and no other agent's. For cross-agent session analytics, agentsview covers the same ground more broadly but without the limit diagnosis.

The logs are local and undocumented. tare sees only the sessions on the machine it runs on, and the README does not address usage from another machine or another client drawing on the same plan, which a limit diagnosis would need to count. It also depends on a log format Anthropic can change in any release. The open directions listed in the README include "better token estimation for tool results", so the per-tool attribution is at least partly estimated.

The repository belongs to Kelviq, whose homepage describes a merchant-of-record and usage-billing platform for SaaS and AI agents, and the README links to a Kelviq blog post. The tool asks for no account and has no upsell, but it is content from a company that sells usage metering.

It is on the watchlist: one contributor, a month old, and parsing a format its author does not control.

Repo: https://github.com/kelviq/tare (MIT). The README gives no star count. The GitHub API on 2026-09-14 showed 267 stars, 13 forks, 21 commits from a single contributor, creation on 2026-08-12 and a last push on 2026-08-28.