Use a local model from a compatible client
LlamaBox includes an optional authenticated HTTP API. It is disabled by default and listens on this phone unless you enable local-network access.
Start locally
- Load a compatible local model in LlamaBox.
- Open the API server section in Settings and enable it.
- Use the displayed address and generated bearer token in your compatible client.
- Keep local-network access off when only this phone needs to connect.
Supported routes
GET /v1/modelsGET /v1/models/{id}POST /v1/chat/completions, including streaming responses.
This is a compatible subset, not every endpoint or feature of the OpenAI API. Developer-only admin routes are not part of the normal public capability set.
Example for a local client
Replace the placeholders with your displayed model ID and token. The configured default port is 3847; use the actual address shown in the app.
curl http://127.0.0.1:3847/v1/chat/completions -H 'Authorization: Bearer YOUR_TOKEN' -H 'Content-Type: application/json' -d '{"model":"YOUR_MODEL_ID","messages":[{"role":"user","content":"Draft a short greeting."}]}'Local-network access is a separate choice
Enabling LAN access allows authorized clients on your network to submit prompts and receive answers. Traffic uses HTTP, not encrypted HTTPS. Keep the token private and use a network you trust. Inference remains on the phone, but replies sent to another client leave it.
Background and busy state
The app and loaded model must remain available. Android can suspend background work. Generation shares the model with in-app chat and can report a busy state; this is not an always-on hosted service.