A second pair of eyes for a lossy video feed.

Cradle is a macOS overlay that reads a No Man's Sky stream and tells me what is on the screen. The console sits across the room, the save file is encrypted, and the game has no API. The only way in is the one I have: pixels.

The technique is bigger than the game. Most screens worth reading have no API behind them: a trading terminal, a devops dashboard, anything that shows state and takes no input. If a person can read it, a program that reads pixels can read it too, and it never has to touch the thing it watches. Cradle is read-only and click-through. It cannot press a button, which is the point.

Compressed video is a bad witness. Remote Play eats small glyphs, OCR drops the first character of a row often enough to matter, and two inventory slots can flatten into one line. My first catalog came from clean wiki screenshots and matched about 89 of 1,063 observations on the live feed, roughly one in twelve. I keep a log of what broke and what each fix measured, because that number is the project.

Under the reading sits a deterministic engine: 125 trade goods, a quote log across roughly 57 stations and about 2,000 goods rows. Given what I hold, it computes which stations to hop next, and every recommendation can show its arithmetic. No ML model anywhere. The same scoring runs twice, in TypeScript behind the review desk and in Swift inside the overlay, with about 285 tests on the Swift side so a fix on one cannot quietly change the other.

It fingerprints its own icons from my labelled trade screens instead of using hand-drawn art. The limit is real: it runs on one machine, plays one game, and I have not measured the pipeline on anything else. But nothing in it is specific to No Man's Sky. Swap the catalog and the icon lexicon and the same overlay reads a different screen.