Remote LAN inference
Use a PC on your Wi‑Fi to run large / GGUF models and stream answers into Nova. On-device LiteRT remains the default. On-device GGUF is not supported in the same APK.
Host (llama-server)
llama-server -m /path/to/model.gguf --host 0.0.0.0 --port 8080
Note your PC’s LAN IP (e.g. 192.168.1.42). Keep the host on a private network;
use a firewall and an optional API token.
Nova client
- Settings → Remote LAN inference
- Backend → Remote LAN (persists immediately)
- Base URL →
http://192.168.1.42:8080 - Model id → whatever your server expects (often the GGUF name or
local-model) - Optional API token
- Test connection, then Save (URL fields also save when you leave the screen)
When Remote LAN is active:
- Chat streams from the LAN host (composer shows Remote LAN)
- Local LiteRT is not loaded for chat (pre-warm is skipped)
- The “No AI model installed” banner is hidden
- Android allows cleartext
http://to LAN hosts via network security config
Deterministic device shortcuts (alarms / open app) still run on the phone; full tool-calling via the remote model is not enabled in v1.
Security
- Trusted private Wi‑Fi only
- Do not port-forward the host to the public internet
- API tokens are stored in secure prefs and are not included in settings backup
Related
- Models — on-device catalog
- Plan:
docs/superpowers/plans/2026-07-17-lan-remote-inference.md(repo only; not published)