The PureGym app only shows how many people are in the building right now. It doesn't tell you how busy a typical time will be. I wanted to train at quiet hours without guessing, so I had one of my Hermes profiles, Uhura, build that history herself. She runs on a VPS at home, and the gym became one more skill with four scheduled jobs.

1. Sign in with a POST to auth.puregym.com/connect/token using grant_type=password, my email, my 8-digit PIN, the scope pgcapi offline_access, and an iPhone user agent. The bearer token lasts 60 days, is cached with 0600 permissions, and re-authenticates on a 401. The PIN is read at runtime from a self-hosted secret manager and never logged.
2. Probe before building. /busyness, /forecast, /popularTimes and /history all returned 404, and every v1 route returned 403. Only /api/v2/gymSessions/gym?gymId=465 worked, giving TotalPeopleInGym and AttendanceTime, so the history had to be collected myself.
3. Sample that endpoint every 20 minutes from 05:00 to 23:00 with cron, writing to SQLite. In the first eight days it made 442 attempts and 277 returned a number. Null counters are logged as misses, because a missing reading is not zero people.
4. Score two-hour windows from 06:00 to 22:00. Each gets an index and a confidence label: measured, emerging, early or modelled. With only two samples the brief says "early signal" instead of faking a forecast. At 07:00 she posts the three quietest windows, the live count, and the one to avoid.
5. Nudge only on rest days. At 21:05 she sends the current headcount, and stays silent if the door record already shows a visit. Checking that record against my session log showed an open visit reports Duration as minutes to midnight, and lockers account for the 5 to 10 minute gap.

After eight days, a weekday afternoon runs at 11 to 15 people, while Tuesday 18:00 averaged 47 and peaked at 51. The same data feeds a 7 by 19 heatmap behind a tailnet-only URL.