Measure whether your AI agents save time and improve quality
Project Eval reads what your agents already record, works out tasks, time, cost, friction and quality, and labels every number with its source: measured, read from logs, inferred by a stated rule, or estimated by a person.
Start small, go further when it earns it
Each step uses the same rules and the same report, so numbers stay comparable as you move up.
- Step 1Evaluate now
Browser
Drop your Claude Code or Codex session logs into this site and get a report. Nothing is installed and nothing is uploaded.
- Step 2Set up locally
Local
Run the collector with Docker Compose and point Claude Code's telemetry at it. Live numbers, kept on your machine.
- Step 3Deploy for a team
Team
Install with Docker Compose or Helm into your own cluster. A default-deny egress policy keeps the data inside it.
Every number says where it came from
A count read from a log is not the same as a time saving someone estimated. The report keeps them apart. Hours saved appear only when you give a baseline, and sampled numbers carry a 95% range.
Hover or tap a label to see its rule.
Illustrative values.
Open source and self-hosted
FastAPI, Postgres, Redis Streams and a LangGraph grading worker. Your data stays where you run it.