LOCAL CODEX RUNTIME · OPEN METHOD
Your day has 24 hours.
Your agents can have more.
Measure the work happening in parallel. A local, inspectable score from completed Codex turns, averaged over 30 full days.
uvx codex-agent-hour-tracker --share --format json
Requires uv and local Codex sessions. Runs on your machine.
Using pip instead?
python -m pip install codex-agent-hour-tracker
agent-hours
--share --format json
Python 3.11 or later. Run the score command in your terminal, then paste its JSON below.
One hour on the clock.
Three agent-hours.
Runtime adds across agents. Three agents each working for one hour count as three agent-hours, even when they run together.
This measures completed agent runtime. It does not measure output quality, productivity, or human time saved.
Each full lane represents one agent-hour. Motion illustrates concurrency.
How much runtime
per human hour?
Explore the ratio of agent runtime to the time you put in, using the same 30-day window.
ESTIMATED AGENT LEVERAGE
2.17×
2.17 agent-hours for each human hour.
Human hours = weekly hours × 30 ÷ 7. This is a scenario, not a measured archive score. It never changes your share card.
Your score.
Your machine.
Paste the public JSON from the command above. Your browser validates the aggregate fields and renders the card locally.
Showing fabricated example data.
Only versioned public score JSON is accepted. Private archives, extra fields, IDs, paths, and daily logs are rejected. Nothing is uploaded or saved by this page. The local tracker reads timing metadata without decoding conversation content. Review your card before sharing.
Card format
1200 × 630 · Example is labeled on the image.
30 completed days
Yesterday and the 29 days before it, in your selected timezone. Zero-use days stay in the average.
Completed agent runtime
Active timing from local Codex sessions. Concurrent turns add together; matching turn IDs are deduplicated.
Scope stays visible
Interactive-only is the default. Add
--include-exec to count batch / exec sessions too.