How We Built Offlex: An Offline AI Chat App for Android
By the AI Segment team · Published 2 November 2026
Most AI apps send your words to a server. We wanted to find out how far we could go with none at all. This is how we approached Offlex, our free on-device AI chat app.
The goal: AI with no servers
We started with one rule: your data stays on your device. That rule shaped everything else. Offlex has no backend servers, no accounts, no analytics and no trackers, so there is no server-side copy of anyone’s conversations to protect, because that data is never collected in the first place.
The stack
Offlex is built with widely used open-source tools, so we could focus on the product:
- Expo and React Native for the Android app.
- llama.rn to run language models on the device.
- A local SQLite database for chat history.
- Public repositories such as Hugging Face as the source of open-source models.
Letting users choose their model
Rather than bundling one model into the app, Offlex lets people choose and download the models they want. Model files can be large, so this also lets people manage their own storage and delete models when they no longer need them. The download is the only thing that requires the internet.
Permissions as a promise
A privacy claim is only credible when the details back it up. Our privacy policy lists each Android permission and why it is requested. Internet access is used to fetch model files, photo access lets people pick images for vision-capable models, and those images are processed locally, never uploaded.
What this means for client projects
Building Offlex taught us the practical trade-offs of on-device AI, including storage management and the need to design around what a phone can handle. On-device AI is a good match when data must stay private or when connectivity is unreliable, and cloud models remain the better option for many other tasks. Knowing when to use each one is the useful skill.
