Free
Gauge
$0 forever $0 forever
See where every token goes. Local audit, no card needed.
- Local transcript audit (in-app + CLI)
- Category breakdown & waste findings
- Estimated savings teaser
- MIT-licensed hypermile-audit CLI
Usage copilot for Claude Code and Codex · Windows, macOS, and Linux
Stop hitting your Claude Code
weekly limit.
Hypermile shows what burns your tokens across Claude Code and Codex, forecasts the hour you will hit your cap, and strips the noise from every prompt so the plan you already pay for goes further.
No account. No card. The free audit runs locally and sends nothing off your machine.
~50%
Approx. tokens saved
Works in
-
Claude Code -
Codex
Same plan.
More miles.
Outcome over upgrade prompts
The limit resets weekly. Deadlines do not. Four ways Hypermile stretches the plan you already pay for.
-
See the burn before it blindsides you
Live breakdown per project, session, and tool. Know which agent or repo is eating the week instead of guessing after the lockout.
-
Forecast the exact hour you will hit the cap
Burn rate projection names the day and hour. Alerts at 60%, 80%, and 95% give you time to slow down or let compression catch up.
-
Cut waste without leaving your machine
Reversible local compression strips logs, JSON, and repeat reads. Originals stay one click away. Nothing uploads.
-
Built for Windows and WSL first
Native paths, config backups, and clean uninstall. Mac only tools leave this crowd capped by Wednesday with no good option.
Three steps, same subscription
Hypermile sits between your coding agent and the API, reading, forecasting, and trimming, without changing how you work.
- 1
See the burn
A live breakdown per project, per session, per tool. No more guessing which agent or which repo is eating the week.
- 2
Forecast the cap
Burn rate projection tells you the day and hour you will hit your limit, with alerts at 60, 80, and 95 percent.
- 3
Cut the waste
Reversible local compression strips logs, JSON, and repeat reads. Originals stay one click away. Nothing leaves your machine.
Fewer tokens. Same answers.
Compression on noisy tool output, measured by the local replay harness. Source files are left alone. Mixed sessions save less; run the free audit to see your mix.
Token savings by scenario
-
Code search (100 JSON results)
90% saved -
CI log (cargo test spam)
99% saved -
GitHub issue (80 comments)
88% saved -
PowerShell listing (Get-ChildItem recurse)
83% saved -
Multi-tool agent (JSON + log + HTML in one turn)
95% saved
What stayed intact
Fewer tokens does not mean fewer answers. Errors stay. Buried facts stay retrievable. Source files never get rewritten.
-
4/4
JSON retrieval
Error, code, resolution, and affected count at log index 67 kept inline.
-
100%
HTML article recall
Every article word survived extraction. Scripts and styles were dropped.
-
3/3
Needles recovered
Facts buried in compacted JSON, logs, and HTML either stay inline or resolve from the local originals store. The model can pull any stub back.
-
Stable
Prompt-cache prefix
Already-sent compressed blocks replay the same bytes on the next turn.
-
Untouched
Source files
Code passes through. We compact tool noise, not your repo.
Replay harness on synthetic tool output (5 noisy scenarios + code passthrough). Not production telemetry. Full results and method · Run the free audit on your own transcripts.
Start with the free audit
One command reads your local Claude Code transcripts and shows where the tokens went. No signup, no upload, MIT licensed. It is the fastest way to see how much of your week is noise. Full install, privacy, and sample report on the audit CLI page.
$ cargo install hypermile-audit $ hypermile-audit Prefer Node? npx hypermile-audit
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 for Windows: https://hypermile.dev/download
The differentiator
Never get surprised by the cap
Hypermile watches your rolling 5 hour and weekly windows and projects the exact moment you will run out. You get a heads up while there is still time to slow down, batch work, or let compression pick up the slack, not a dead terminal at 4 pm.
- Alerts at 60%, 80%, and 95% of your window
- Burn rate projection updates as you work
- Lives in the Windows tray. No dashboard to babysit
Hypermile now
At this pace you'll hit your weekly cap Wed 2:14 pm.
Burn rate 1.8× your 7-day average. Compression is on — turn up adaptive mode to buy ~2 more days.
Sample tray notification · times illustrative
Two plans. No account snooping
The free audit stays free. Compression, forecast, and the inspector are a flat price — and we never read your Claude account to decide what you pay.
For daily drivers
Pro
$7 /mo $5 /mo, billed yearly
Everything in Gauge, plus the engine that stretches your plan.
- Live dashboard & weekly-cap forecast
- Reversible local compression
- Before/after request inspector
- Adaptive mode as you near the cap
Founder pricing: the first 300 seats get 40% off Pro for their first 12 months — around $3/mo on the annual plan, then list. Paid licenses are final; cancel anytime. Refund policy.
Frequently asked questions
Does any of my code or data leave my machine?
No. Hypermile runs entirely on your own device. Compression, the originals store, and the dashboard are all local — nothing is uploaded, and there's no account required to run the audit or see your usage.
Does it work with Claude Code and Codex?
Both are first-class. Hypermile wires up Claude Code and OpenAI Codex on native Windows and inside WSL2 automatically, with timestamped config backups and a clean uninstall that reverses every change.
Does it work with Cursor?
Partially, and we won't pretend otherwise. Cursor routes most of its AI traffic through its own cloud, which no local proxy can intercept — so full request compression isn't possible there. What Hypermile can do for Cursor is compress command output and support custom-API-key setups where you point Cursor at your own key. For Claude Code and Codex, the full engine applies.
When can I buy it?
Hypermile launches as a finished product — the desktop app, this site, and the audit CLI ship together, with no sign-up gate and no email capture. The free audit CLI already works today; grab it from the download page. Paid plans (Pro and Max) unlock the compression engine, inspector, and adaptive mode.
Won't compressing prompts make the model dumber?
It shouldn't, and the replay harness is how we check. Compression is reversible: originals stay on your machine and the model can pull any of them back when it needs the full detail. The measured savings and quality checks live on the Benchmarks page.
Do you have a database, or track me?
No. This website is a static set of pages — no accounts and no forms. The only measurement is our own self-hosted, cookieless page-view counter: it sets no cookies and never stores your IP address, which is why there's still no cookie banner. The app itself is local-first: your usage data lives on your machine, not ours.
What about macOS and Linux?
They're on the roadmap after the Windows and WSL release. The app is built on a cross-platform stack, so the ports are mostly a matter of wiring and testing. Windows and WSL come first because that's the crowd with no good options today.
What happens if the proxy crashes mid-session?
It fails open. If the proxy dies mid-session, the wiring is removed and your coding agent keeps talking to the API directly as if nothing was installed. You never get locked out of your own tools because of us.
How does pricing and the free tier work?
The Gauge tier — the local transcript audit in the app and the MIT hypermile-audit CLI — is free forever. Live forecast, compression, the inspector, and adaptive mode are a flat price: Pro or Max, one price regardless of which Claude or Codex plan you're on. The first 300 seats get 40% off their first 12 months.
Do you offer refunds?
No. Pro is a low-price digital license issued as soon as you pay, and Gauge is free so you can try before you buy. Cancel anytime to stop the next charge; unused time is not refunded. We will reverse a charge if we failed to deliver a key, billed you twice, or the law requires it. Full policy: Legal — refunds.
Stretch the plan you already pay for
Grab Hypermile for Windows. No account. No card. Cancel anytime on paid tiers. The free audit stays free.