Privacy & data

crowd-cast is built so that you decide what gets recorded. This page is the single place that spells out what’s captured, what isn’t, and how to keep sensitive things out.

What is captured

• Screen video of the applications you selected, while they’re in the foreground.
• Keyboard and mouse input (keys, mouse movement, clicks, scroll), but only while a selected app is in the foreground.
• Basic metadata such as which app is focused and your display resolution.

What is not captured

• Anything from apps you didn’t select. When a non-selected app is in front, the video is a black screen and input logging stops.
• Audio: it’s off by default.
• Anything while you’re idle. crowd-cast pauses after 2 minutes without activity and records nothing until you return. See idle detection.

How to keep things private

The model is an allowlist: only the apps you explicitly choose are ever recorded. A few ways to use that:

• Use a separate app for sensitive work. The most reliable approach is to keep private activity in an app that simply isn’t on the list.
• Add only what you’re comfortable sharing. The more apps you record, the more useful the data, but it’s entirely your call which ones.
• Use the panic button. If something sensitive slips into a recording, hit Delete last 10 minutes in the tray to discard recent, not-yet-uploaded footage.

How we use it ourselves

We record one browser (e.g. Firefox) for research, a text editor (e.g. Cursor), a terminal (e.g. Ghostty) and a PDF viewer (e.g. Preview). We use a separate browser (e.g. Chrome) for sensitive things like email, messaging and contracts. That browser simply isn’t on the list, so it’s never recorded. When we switch to it, crowd-cast captures a black screen and stops logging input. We’ve found this covers most of the working day while keeping anything private completely out of the recording.

Where your data goes

Finished segments upload to p(doom)‘s research backend over secure, signed URLs, and (when you’re signed in) are attributed to your account. Local copies are removed once they’ve uploaded. To understand how the dataset is used, read the research write-up: crowd-cast: Crowd-Sourcing Months-Long Trajectories of Human Computer Work.