Kenkage API

Kenkage API turns your phone into a shared, on-device AI runtime: it runs real models (Gemma 4 / Gemma 3 for chat, Kokoro for text-to-speech) locally and exposes them through a local, OpenAI-compatible REST API — so any app on the same device, LAN, or the internet (via your own tunnel) can call a real AI API without a cloud key, and without every app bundling its own model runtime.

Point any OpenAI SDK at it — openai-python, openai-node, or a raw curl — and it works as long as the app is running and a model is loaded: http://127.0.0.1:8080/v1/...

What's inside:

• Gemma 4 E2B (default chat model, ~2.6GB) and Gemma 3 1B INT4 (lite fallback, ~584MB) via LiteRT-LM
• Kokoro TTS (2 voices: Anna, Jamary) via sherpa-onnx
• /v1/chat/completions (streaming + non-streaming), /v1/completions (legacy), /v1/audio/speech, /v1/models
• An in-app Playground to test the API without leaving the app

Honest about gaps: no speech-to-text, and /v1/embeddings returns a real 501 — this app never fakes a response.

Network exposure is your choice: loopback-only by default, or opt into LAN / LAN + API key from Settings, with a visible on-screen warning whenever the server leaves loopback.

Built for privacy: inference stays on your device unless you choose to expose it further.

• Initial Store release: on-device OpenAI-compatible API (chat, completions, audio/speech, models), in-app Playground, shared Gemma/Kokoro model store, and loopback/LAN/LAN+API-key exposure modes.