Open contribution

Help train the world's strongest Catan engine.

Atlas learns by playing itself, millions of games over. Every game your computer plays makes the engine measurably stronger — and credits your machine on the live workers page.

Why this matters

The best published Catan AI today (HexMachina, 2025) reaches 54% vs a strong alpha-beta baseline. Atlas's goal is to surpass that, in the open, with contributor compute. No company. No paywall.

What your PC does

Plays Catan against itself as fast as one CPU core allows (typically 2–4 games/sec). Each finished game is uploaded as a ~7 KB record. No GPU needed.

Footprint

One CPU core per worker process. ~50 MB RAM. ~100 KB/sec network upstream at peak. You can run multiple workers if you want to dedicate more cores, or just one to barely notice.

Install in under a minute

Choose your OS and how many workers to run. The install snippet updates automatically.

Will appear as your-computer-name-1 through your-computer-name-2 on the live workers page. Leave blank to use your computer’s hostname.

Each worker uses one CPU thread. The default leaves one thread free for your OS.

Each worker opens in its own PowerShell window — you can see logs and close any one to stop it.

Step 1

Open PowerShell

Search 'PowerShell' in the Start menu. No admin needed.

Step 2

Paste this and hit enter — starts all 2 workers (Windows · PowerShell)

New-Item -ItemType Directory -Force -Path .\atlas-worker | Out-Null
cd .\atlas-worker
Invoke-WebRequest "https://downloads.landover.io/start-worker.ps1?t=$([int64](Get-Date -UFormat %s))" -OutFile start-worker.ps1
1..2 | ForEach-Object {
  $i = $_
  Start-Process powershell.exe -ArgumentList @(
    "-NoExit","-ExecutionPolicy","Bypass","-Command",
    "`$env:ATLAS_WORKER_DIR='$PWD\worker-' + $i; `$env:ATLAS_WORKER_NODE_ID='$env:COMPUTERNAME-' + $i; .\start-worker.ps1"
  )
}
Write-Host "✓ Opened 2 worker windows. Close any window to stop that one."
Step 3

That's it!

You should see ⠋ Playing games: N tick up. Your contributions appear on the live workers page within seconds.

FAQ

Is it safe? What does the binary actually do?
Yes. The binary is a .NET console app that plays Catan against itself locally and uploads the game record (just the move history) to atlas-engine.landover.io. It doesn't read your files, doesn't take any input from your machine, and only makes outbound HTTPS calls to the Atlas API.
Will it slow down my computer?
One worker pegs exactly one CPU core. If you have an 8-core machine and run one worker, you're using 12.5% of your CPU and won't notice. If you want to be safe, run only one worker, or use Windows Task Manager to set the process priority to "Below Normal".
How much data does it upload?
About 7 KB per finished game. At ~3 games/sec sustained, that's ~20 KB/sec upstream — under any home internet's budget.
Can I stop it?
Yes, any time. Close the PowerShell/Terminal window, or run Get-Process atlas-worker | Stop-Process (Windows) or pkill -f start-worker.sh; pkill -x atlas-worker (Mac/Linux). Any in-flight batch is flushed before exit.
What's the deal with `apiKey`?
Currently a shared contributor key so the server can rate-limit. It's not secret — anyone running a worker uses the same one. We may switch to per-user keys later if needed.
How do I know my contributions are actually helping?
Each worker self-reports a nodeId (default: your hostname). The live counter on the main page shows total active workers globally; a per-node breakdown is coming soon. Behind the scenes, every game you play adds to the training dataset for the next model version.
What does the AI model look like?
Currently a small neural network (~500 KB ONNX file) with a value head (predicts who wins) and a policy head (predicts good moves). It's wrapped inside an alpha-beta search for actual play. Every few hours we retrain on accumulated data and the workers automatically pick up the new model.
What if I have a GPU and want to do more?
You can run the trainer on your machine, not just the worker. The trainer downloads the worldwide trajectory dataset and produces new model versions. Ping Jay for an admin key and see the trainer bundle.
Questions or stuck? Email legal@landover.io.