I started this project because I find the R1 unusually compelling as a piece of hardware: its small, characterful, and designed specifically for interaction with an AI agent. One thing to say off the bat, this wasn't in any way motivated by frustration with Rabbit’s software or an attempt to “fix” it. I wanted to explore a different question: how much of a custom, persistent voice-agent experience could I build directly on the device?
I had to build the UI from scratch for this to accommodate R1’s 2.88-inch square display. You tap the tomato to talk, see the live transcription in a foreground overlay, and the message is sent automatically after a short pause. The response appears in the same conversation overlay. There is also an ambient lock display, conversation management, an animated home state, and an offline “Hey Pomi” wake word using Picovoice.
The less visible part of the project was making the agent runtime behave like part of the device rather than a development process. I used Hermes by Nous Research - fantastic platform and I use it heavily outside of this. The agent gateway starts during boot, stores its state locally, and is supervised by Android’s service system. Normally that runs on localhost, but I connected the R1 to my private Tailscale network and had a custom tailwind provisioned URL that only other devices (like my Mac) can access.
The repository includes the Android shell, the Hermes runtime configuration, the UI lab I used to iterate on the 240×282 interface, and scripts for building, provisioning, installing, updating, checking, and recovering the device. It targets a compatible rooted Rabbit R1 running Android 13. It does not redistribute Rabbit firmware or any user credentials.
I originally built this only for my own R1, but I cleaned up the project so other R1 owners can reproduce it: https://github.com/sidmohan0/pomi I’ve just started collecting public feedback, so please let me know what you think. I’m particularly interested in the voice interaction model, reducing the time between speaking and receiving a response, and whether the installation process is practical enough for other R1 owners. I’d also love to hear what people would build with this hardware once the original software is no longer the defining constraint.
PS I used Tailscale which made things like OTA updates and accessing the Hermes Dashboard from my computer much less painful, but it isn't required!