HOW IT WORKS

RIGHT NOW

STEP BY STEP

FINDWe read listings, calendars, news, web search, Bluesky and X

READAn AI model turns what we found into happenings

PLACEEach happening goes on the map, if we can find where it is

CHECKPosts on X and Bluesky tell us if it's on, busy or sold out

EDITDuplicates get merged, and follow-up searches get suggested

SHOWEvery change shows up on open pages within seconds

Every five minutes the server wakes up and does a round of work. First it adds the jobs that are due (each source has its own pace), then it runs the waiting jobs, most important first. It stops when there's nothing left to do, when today's budget runs low, or when a new version of the site takes over. Jobs that use a paid AI model only run when the budget has room for them.

A happening is anything worth knowing about, like an exhibition, a festival, a market, a pop-up or a seasonal sight. Each one keeps its sources, dates, places and what posts say about it. When posts about a happening pile up while it's on, it becomes hot, and we check it every 30 minutes. When three happenings are hot at the same time, the whole day becomes a hot day, with a bigger budget.

Besides happenings, we also bring in posts about Tokyo in general (lines, weather, trains, food) and news stories, for the Latest box. Every post gets an English translation, and the original stays underneath.

There's a daily budget for AI models, and a bigger one on hot days. It's spread across the day, so the morning can't use up the evening's share, and a quarter of it is kept aside for hot happenings. Each call sets aside its expected cost before it's sent, and records the real cost after. When there's no room left, paid jobs wait until there is. At night (01:00 to 06:00 Tokyo time) things are quiet, and we only check happenings that are on right now.

THE JOBS

BIKE SHARE

The map can also show Docomo bike share (turn on "Bike share" on the map). Tokyo has about 1,900 stations, and showing all of them is just a mesh of points that doesn't say much. So we only show the few areas where clearly more stations are empty, or full, than usual right now.

The stations come live from ODPT's open data. To know what's usual, we use Max Halford's bike-sharing-history, which has saved the same feed every 12 to 16 minutes since 2024. We took a year of it (September 2025 to September 2026, about 22,000 snapshots) and split Tokyo into hexagons 1 km across. For each one, we worked out how many of its stations are usually empty or full at each hour, on weekdays and on weekends or holidays.

Rain matters too. When it rains in the evening, people stop riding, and the bikes stay where the day left them, so stations near homes stay empty and stations near offices stay full. On the rainy night of November 4, about half of the areas looked unusual. So we also keep what's usual for rainy hours (1 mm of rain or more in the last 3 hours, from Open-Meteo), and while it rains, an area has to look unusual against both.

An area shows up when it has at least 5 stations and at least 2 more of them are empty (or full) than usual, which also has to be far more than on most days (2.5 standard deviations above the average for that hour). We show the 5 strongest, and each one says how many stations are empty or full now and how many usually are. Bike share data is from DOCOMO BIKESHARE, INC. / ODPT (CC BY 4.0).

BUILT ON NETLIFY

Everything runs on Netlify. There's no other hosting, queue or database service involved.