Local AI already solves some of the privacy issues, but in my experience it has a learning curve and often needs decent hardware to work well. PearPie aims to make private AI accessible to everyone, while still allowing the full local AI experience for who wants it.
In PearPie conversations live on your device and sync peer-to-peer (holepunch stack) between devices you own/trust. This means nothing is stored, can be trained on or analysed, or has the possibility of being leaked.
Currently there are three ways to run models in the app:
- Local: Download Qwen/Gemma models directly in the app, or connect to LMStudio/Ollama if you already have models running there. This is completely free, nothing leaves your “network” of devices.
- Delegated inference: Connect to a remote computer you trust over P2P and use its GPU, same connection as your chats sync over. Also completely free, it exposes the same models you would have in the full local path.
- PearPie models: Hosted models on European infrastructure, paid via credits attached to your public key. When you connect, you're assigned a random relay node that behaves the same way as delegated inference with the same privacy guarantees. All requests are ephemeral and the relay is destroyed when you’re done.
The goal is that you never have to think about which path you're on. There's a "simple" track that feels like any other AI app, and an "advanced" track that opens up model downloads, device connections, and remote inference.
You can toggle a "European providers only" filter to drop closed frontier models and stick strictly to EU-owned infrastructure, meaning no (American) hyperscalers are in the path.
In the app itself there are “regular chats” that start without any context (other than the system prompt) each time. There is also a “projects” feature. This is supposed to be a contained workspace. The idea is that you can build context here over a longer time, going back to fetch context of previous work/past decisions. This feature is also a little more like an “agent”, giving it some extra capabilities like interacting with folders, creating documents etc.
Some trade-offs:
- Identities are not recoverable. Lose your device without it being synced elsewhere and your data is gone.
- Some rough edges exist: core experience is solid but expect some friction. I am working on some stability issues around reconnecting especially. Caching is still best-effort. Agentic capabilities are still being improved/developed.
- Local model speed/quality depends on your hardware. I try to make an “educated” guess on what your system can handle, but this approach is rather simplistic.
- Because it is P2P chats can only sync when both devices are online. If this does not happen a device can fall behind. I have therefore also opted not to allow creation of projects/updating of projects on mobile yet, because this might cause strange experiences.
You can download PearPie on the website. It is available for all major platforms (including mobile) and I am giving away free credits so people can experiment and try without downloading models. I'd genuinely love feedback on the architecture, the UX, and where you think I've made the wrong trade-offs or I can improve.
General questions:
- I opted for Europe only infra, being from Europe this made sense to me (easier GDPR compliance), but I wondered if it is a blocker to use it for people outside of Europe?
- You can connect to local AI apps (LMstudio/Ollama) over PearPie's P2P connections, which appears easier to me than the current remote options they offer. Do people have experience with this? Would this be a feature you would use?