How it works
1. You export your usage
From the OpenAI or Anthropic console, download the usage/cost CSV. It contains token counts per day, model and key — split by cache role and service tier. It does not contain the content of your requests, and we never ask for it.
2. It's analyzed in your browser
When you drop the file in, the analysis runs as JavaScript on your own machine. The file is never transmitted to a server. That's the whole privacy model: there is no upload to trust, because there is no upload.
3. Findings, split by confidence
Every finding carries an honest label:
- Provable — measurable from your export alone (caching gaps, deprecated models). These form the defensible floor.
- Flag — real, but the euro figure needs a judgment (which batch work is non-urgent) or a quality eval (which model calls can downgrade).
- Estimate — needs request logs to measure exactly (duplicate calls). Shown as a benchmark range, never folded into the floor.
4. You fix it — with or without us
Each finding lists the exact fix. If you want the precise number and the work done, the deep audit adds request-log analysis and a fix plan. If you'd rather DIY, everything you need is on the report.