RELAIS

Your own relay station for AI

Relais turns an Android phone into a headless, always-on LLM node. It runs a language model directly on the phone's GPU and neural silicon, and serves it over your LAN as a standard OpenAI-compatible API. Point any OpenAI client at the phone and it answers.

Install it on a spare phone, plug it in, put it on a shelf. There is no chat UI to babysit — it auto-starts, survives Doze, recovers from crashes and OOM kills, and draws a few watts. It is inference infrastructure you leave running.


Talk to it

$ curl -k https://192.168.1.42:8443/v1/chat/completions \
    -H "Authorization: Bearer $RELAIS_KEY" \
    -H "Content-Type: application/json" \
    -d '{"model":"gemma-4-E4B-it",
         "messages":[{"role":"user","content":"ping"}]}'

The LAN surface is HTTPS on :8443, with a TLS certificate generated on the device — hence -k, or pin the cert. Plain HTTP on :8080 is bound to loopback only and is not reachable from the network. The bearer key is generated on-device and stored encrypted; the node is found automatically over mDNS.

Endpoints

/v1/chat/completionsOpenAI chat · streaming SSE
/v1/messagesAnthropic-compatible
/generatenative multimodal · text, image, audio
/v1/embeddings · /v1/rerankvectors and reranking
/v1/rag/documents · /v1/rag/queryon-device retrieval
/v1/audio/transcriptions · /speechspeech in and out
/v1/images/generationsimage generation
/v1/models · /v1/batch · /v1/sessionscatalog, batching, state
/health · /metricsliveness · Prometheus

What it is not

Relais will not outrun a desktop GPU on raw tokens per second, and does not pretend to. Its value is different: always-on, low-power, private, zero-marginal-cost inference. Use it for small-model workloads, agents, automation, and home-lab pipelines that need a local endpoint that is simply always there.

Sovereignty

Requirements

Android12 or newer
Deviceone LiteRT-LM supports on GPU
Storagethe default model is a ~3.7 GB download
LicenseAGPL-3.0 · forked from google-ai-edge/gallery