Free tool · Local only
See where your tokens go. Nothing leaves your machine.
hypermile-audit is the free CLI behind Hypermile’s Gauge tier. One command reads your local Claude Code and Codex transcripts and prints a plain-English breakdown of the week — categories, waste, estimated savings range. No account, no upload, MIT licensed. As of August 2026 this is the fastest way to measure before you change anything.
Install
Two advertised commands. Prefer Node for a one-shot run; prefer Cargo if you already have a Rust toolchain.
$ npx hypermile-audit $ cargo install hypermile-audit $ hypermile-audit
Source and releases:
GitHub.
npx may download the binary once from GitHub Releases; the
audit run itself makes no network calls.
What it reports
Point it at your local transcript directory (Claude Code’s
~/.claude by default, including inside WSL2) and it scans
recent sessions. The terminal report covers:
- Totals — context processed, cache-read share, session and project counts for the window (last 7 days by default).
- Where it went — a category split such as bash/tool output, file reads, search, model output, and your prompts.
- Waste findings — repeated file reads, compressible JSON or log output, huge single tool results, and sub-agent share when those markers are present.
- An estimated savings range — from those detectors, not from fleet telemetry. Treat it as a planning number for your transcripts.
Useful flags: --days 30, --project,
--json, --html for a shareable file, and
--redact-paths if you post a report. Full flag list is in
--help.
hypermile-audit — report
HYPERMILE AUDIT — last 7 days, 3 projects, 41 sessions Total context processed ..... 182.4M tokens (94.1M cache reads) Where it went: ████░░░░░░ 38% Bash output 69.9M ███░░░░░░░ 27% File reads 49.2M ██░░░░░░░░ 18% Model output 32.8M █░░░░░░░░░ 9% Your prompts 16.4M █░░░░░░░░░ 8% Search + MCP 14.1M Waste found: • 12.1M tokens re-reading the same 87 files (top: src/big-module.ts x9) • 21.3M tokens in compressible JSON / log output • Top offender: `npm test` output, 4.2M tokens across 31 runs Estimated savings with Hypermile: 41-56% (~2.3 extra days before your cap) -> Download: https://hypermile.dev/download
Same layout as the homepage demo (spec 01 sample). Not production telemetry, not a promise of savings on your workload.
Privacy: local-only
The audit reads *.jsonl transcripts on disk. It does not
upload them, and it does not send telemetry. The printed report uses
file names, tool names, and aggregate token counts — not file
contents or prompt text. If you share an HTML or JSON dump, add
--redact-paths so directory layout is hashed.
That is the same local-first rule as the rest of Hypermile: Gauge keeps the in-app audit on your machine; Pro compression also stays local. Cursor remains partial for the desktop app (most Cursor traffic is not interceptable by a local proxy) — the CLI only reads transcripts that already exist on disk.
Then stretch the week
The CLI is the measurement step. Hypermile Gauge — free — is that audit in the desktop app. Pro adds a cap forecast and reversible compression so the noisy parts of each prompt never reach the model. No fabricated fleet averages here; harness results live on Benchmarks.
Related
- Claude Code usage limits and the Claude weekly limit
- Codex usage and the Codex weekly limit
- Claude Code WSL2 setup if transcripts live inside a distro
- Token calculator for a paste-in heuristic (API dollars, not subscription windows)
Frequently asked questions
Does hypermile-audit upload my transcripts?
No. The audit reads local JSONL transcript files on your machine and prints a report. It does not upload prompts, file contents, or usage data. The npx installer may fetch the binary once from GitHub Releases; the audit run itself stays offline.
What does hypermile-audit report?
A category breakdown of where context went (tool output, file reads, prompts, model output), waste findings such as repeated file reads and compressible logs, and an estimated savings range from those detectors. Your numbers come from your own transcripts.
How do I install hypermile-audit?
Run npx hypermile-audit for a one-shot install, or cargo install hypermile-audit if you have a Rust toolchain. Both commands are the ones advertised on this site. Source is MIT on GitHub.
Is the CLI the same as Hypermile Gauge?
Gauge is the free product tier: the in-app local audit plus this MIT CLI. The CLI is the no-app starting point. The desktop app adds the same local audit in a UI, then Pro unlocks forecast, compression, and the inspector.
Does hypermile-audit work on Windows and WSL2?
Yes. Run it on native Windows, macOS, or Linux, or inside a WSL2 distro so it reads ~/.claude where WSL2 Claude Code actually writes. It does not need a Hypermile account.