Touch Grass

Touch Grass answers one small question: which two hours tomorrow are actually worth going outside for? You run one Python file on your own laptop, it pulls about 70 days of hourly weather for your city from Open-Meteo, and it hands you a single offline web page with three ranked two-hour windows, a temperature chart, and a plain table of every daylight hour.

The part that makes it worth trusting is that it shows its own homework. Before it forecasts tomorrow, it re-runs the same pipeline on the last three fully observed days and prints how far off it was, right next to the simplest baseline there is: the same hour two days earlier. On the Paris example week it lost on temperature but won on wind and cloud, and the page says so in plain numbers instead of hiding it.

There is nothing to sign up for and nothing in the cloud. No account, no API key, no GPU: the open-weight TabPFN model runs on a laptop CPU in about five minutes, and the saved weather file lets you re-run it fully offline. The project is open source and has no GitHub stars yet.

Honest limits: this is a single-file hackathon project, not a weather service. The hosted page is the example built on 8 October for 9 October in Paris, not a live forecast for today; to get fresh windows for your own city you run the script yourself, and the full run takes five to six minutes on a laptop CPU.