How the audit works
Route reads the logs your AI tools already keep on your own machine. Nothing is uploaded, and it needs no API key. Press play to watch a real run finish in a few seconds.
You run one command
No install, no account, no key.
It reads your own logs
Every agent role, project and request you already paid for, priced from raw tokens.
It names the change
Only for jobs it has tested. It stays quiet on the rest.
A real run, replayed
This is a recording, not a live scan. A web page cannot read files on your computer, and we would not want it to. Every line above came from a real route audit on the author’s machine. For your own numbers, run the same command — the data never leaves your laptop.
What is built, and what is not
Everything marked Built runs today and produced the numbers on this site. Everything marked Coming soon is honest about not existing yet.
Read the estate
Agent roles, skills, repos and every request across every project, priced from raw tokens.
Produced the $690 and the overspend figure on the dashboard.
Score every model
All 19 Token Factory text models on the same tasks, one blind judge, cheapest that clears your bar wins.
It overruled our own hand-picked model.
Route the calls
OpenAI-compatible endpoint. Cursor, Cline and Continue switch automatically; every call leaves a costed receipt.
Works today for anything speaking the OpenAI protocol.
Scan from the browser
Drag your ~/.claude/projects folder onto the page and see your own numbers, parsed entirely in your browser.
Nothing would be uploaded. We would rather ship a real recording than a fake scan.
Apply the change for you
One click writes the model change into your agent files, with a diff to review and a one-command undo.
Today Route prints the change and you run it. Editing your files needs to be something you trust first.
Claude Code support
An Anthropic-protocol adapter so Claude Code routes through Route the way Cursor already does.
Claude Code speaks a different wire format. Until then we hand those users a one-line config change.
Automatic upgrading
Answer on the cheap model, detect a weak answer, pay for the strong one only then.
We built two versions and measured both. Neither worked: a small model is wrong confidently and in good form, so a cheap confidence check cannot see it. Published as a negative result.
Spot repeat work
Connect your repos and surface the tasks you keep paying an agent to redo, ranked by what they cost you.
Today the audit attributes spend per project. Detecting repetition is the next layer.
Team control plane
Shared policies, spend caps, per-project ledgers and approvals for whoever signs off on the bill.
This is the commercial product. The audit stays free.
Artwork: da Vinci, Claesz, and the Met’s open-access collection — all public domain or CC0, credits in art/CREDITS.md. Same plates the AX dashboard uses.