LlamaBox network behavior | What is offline and what is not

Answer-first summary. Chat generation in LlamaBox does not require a network connection once a compatible model is stored on the device, and there is no LlamaBox inference server. Network access is used when you explicitly download a model, when the local-network API is active, and — separately from the app entirely — when you use this website or support email.

What works with the network switched off

Loading a model already stored on the device.
Chat generation, including streaming tokens.
Reading and writing local conversation history.
Vision analysis with a compatible model and matching projector already on the device.
Generation settings and model switching.

What requires a network connection

Activity	Leaves device	User-initiated	What the other party receives
Local chat generation	No	Yes	Nothing — no server involved
Image analysis	No	Yes	Nothing — processed on-device
Local history read/write	No	Yes	Nothing — local SQLite
Model download	Yes	Yes	Repository receives standard HTTP request metadata
Local-network API (when active)	Local network only	Yes	Clients you point at it, on your own network
App analytics / telemetry	Not implemented	—	No such endpoint exists
Crash reporting	Not implemented	—	No such SDK is present
In-app update check	Not implemented	—	Updates come via Google Play
Website visit	Yes	Yes	Static host logs; Google Fonts request
Support email	Yes	Yes	Your email provider and the project email inbox

Verify it yourself

Behavioural verification beats reading a policy. On your own device:

Download or import a compatible GGUF model and confirm it loads.
Send a prompt and confirm you receive a generated response.
Enable airplane mode, or disable Wi-Fi and mobile data.
Send another prompt. Generation should proceed normally on-device.
Optionally inspect Android's per-app data usage, or route the device through a network monitor, to confirm no traffic accompanies generation.

If step 4 fails, the offline claim is wrong for your configuration and we want to hear about it: work.aalhad@gmail.com.

Precise statements about model downloads

Model downloads are always initiated by you, never in the background on first launch.
The repository hosting the file receives the ordinary metadata of an HTTP request — IP address, requested path, user agent. This is inherent to fetching a file over the internet, not something LlamaBox adds.
LlamaBox does not intentionally attach chats or prompts to model-download requests.
To avoid all network activity, transfer a GGUF file to the device yourself and import it.

Text-to-speech caveat

Voice output uses the Android system TTS engine you have selected. Some engines synthesise on-device and some send text to their vendor's cloud. That is a property of the engine you chose in Android settings, not of LlamaBox. If offline voice matters to you, verify your selected engine's behaviour.

How these claims were established

Verified by reading the application source tree and dependency manifest:

No analytics, product-telemetry or crash-reporting SDK is declared as a dependency.
Model discovery and downloads contact Hugging Face when initiated from the app.
No LlamaBox-operated API endpoint is referenced anywhere in the app.
Conversations are persisted through local SQLite.
The local-network API, when active, serves compatible clients while inference remains on-device.
Declared Android permissions are: internet, external storage read/write, record audio, system alert window, and vibrate.

Not independently verifiable by a visitor today: the Android app source is not yet public, so you cannot compare a shipped binary against these statements yourself. They are declared product behaviour, testable behaviourally via the airplane-mode procedure above. Public source release under AGPL-3.0 is planned; see release status.