This week at Parley Labs Offline-first isn't really a feature. It's a statement about who owns the compute. We encourage you to install our PWA for a better desktop experience. #heyparley #AI
Most "AI apps" are just a browser tab pointed at someone else's GPU. Close the lid on a plane and you've got nothing. This week Parley started working with no internet at all. Install it like a native app. Go offline. Keep generating — the on-device models run on your own GPU, so there's no network round trip and nothing leaves your machine. The hard part wasn't offline mode. It was the accounting. If it works offline, what stops usage from being infinite? → Offline generations meter against a local ledger — your balance visibly drops → At zero, generation stops until you top up → The moment you reconnect, the queue settles itself The piece I'm most pleased with: settlement is idempotent. Every offline generation carries its own ID, and the receipt is written before the charge. Flaky reconnect, laptop closed mid-sync, duplicate network event — you cannot be double-charged. The worst case is that we under-charge you. That's the correct direction for that bug to fall. Two things we refused to compromise on: 1. Offline can't strand you. The app always prefers the network when it exists and only falls back to cache when it doesn't. A cached shell can never trap you on a dead version of the app. 2. Your downloaded models are sacred. Cache cleanup is scoped so it can never wipe the model weights you already pulled down — because recovering from that needs exactly the thing you don't have. One honest caveat: you need to be online once, to install the app and download the model. After that you're free. Offline-first isn't really a feature. It's a statement about who owns the compute.
