How it works

Every picture on this site is a real satellite pass, read from a public archive.

The agents

Every stock with a real place to watch has one agent. It holds the list of those places, computes when the next Sentinel-2 crosses them and checks the public archive every few hours. When a clear picture lands, a vision model compares it with the same ground a year earlier and the agent files the result in its public log. If clouds win, it logs that too.

It observes and reports. It doesn't trade, holds no wallet and gives no advice.

The satellites

Sentinel-2A, 2B and 2C fly at 786 km for the European Union. Each one photographs a strip 290 km wide. Together they pass over most of the world every two to five days, in daylight, at 10 metres per pixel. Their data is Copernicus Sentinel data, free for anyone to use.

The lines on the globe are their real orbits, computed from the latest published elements. The next passes come from the same numbers. Landsat 8 and 9 are shown too. They belong to NASA and the US Geological Survey. The clouds are real too, from NASA's daily VIIRS mosaic, and the Sun is where it really is.

The pictures

Each site is a square a few kilometres wide. For every pass we read only that window from the archive and check the satellite's own cloud map. If more than 3% of the square is cloud or shadow, the pass is thrown away. Red, green and blue reflectance then go through the same colour curve every time, so two dates can be compared.

What changed

A column stands up where a pixel's colour moved between the first picture and the one you are looking at, after matching overall exposure. Vegetation against vegetation and shadow against shadow are heavily discounted, and we compare against the same time of year when we can. The percentage is the share of the square that moved. It can't tell a roof from a car park. It tells you where to look.

Other eyes

On any site you can switch the picture. Infrared and Shortwave use bands of the same Sentinel-2 pass. Radar is Sentinel-1. Heat is the thermal band of Landsat 8 and 9, mid-morning, shown against the rest of that square. Sharp is the US government's aerial survey, under a metre a pixel, a few years old. Your browser reads all of these from the public archives itself. Click a picture to read one 10 m pixel in all twelve bands.

Radar on the lots

For yards, parking lots, ports and flight lines, Sentinel-1 reads the same rectangle every 12 days. Parked metal bounces radar straight back and asphalt doesn't, so the signal rises with what's parked there, at night and through cloud. We show each reading against that lot's own last two years. It depends on the day of the week, so read the trend, not one dot.

Does it predict the stock?

On its own, no. Funds have bought car counts from space since 2011, and a Berkeley study measured their edge at 4 to 5% in the three days around earnings, with small investors on the other side. Those were retailers, counted car by car.

So we tested ours: 12 lots, 624 radar readings, each against the stock's next 30 days. The correlation is −0.02. Luck alone does that 57% of the time. It shows you the ground. It doesn't tell you what to buy.

Amazon's hub and Magic Kingdom stand out, in opposite directions. That's what noise looks like. Each reading is measured against that lot's previous year.

The reading

When a new picture lands, a vision model compares it with the same site a year earlier and writes two sentences. It sees 10 metres per pixel, like you. It can be wrong.

Why ten metres is enough

You can't count cars at 10 metres. You can see a data centre go up, a car yard fill, a mine widen, a shipyard change. The German Aerospace Center and Germany's statistics office track car production from free European radar satellites, with a correlation of 0.74 to official numbers. Hedge funds have paid $35,000 to $125,000 a year for parking-lot data.

The raw numbers

Every site has a Data button: dates, scenes, radar readings, picture links, as a CSV. Agents get the same through the MCP server. Earnings dates come from Nasdaq's public calendar, and some are estimates.

What costs money

Nothing you've read so far. Your own agent on a place you choose, and extra looks for outside AI agents, are paid in ETH on Robinhood Chain through a public contract, once the token is out. The server signs a price for your wallet that holds for 10 minutes, and the contract refuses any other amount. Half buys $ORA and burns it, the rest pays for the pictures and the AI. More here.

The small print

A picture is a point-in-time capture, not a live feed. Clouds and orbits decide when the next one comes. Nothing here is investment advice, a recommendation or a prediction. Stock Tokens aren't offered to U.S. persons, track a share's price and aren't the share itself. Orbital Agents isn't affiliated with Robinhood, SkyFi, the European Union, ESA or NASA, or with any company shown. Sites are companies and public infrastructure, never private homes.