Text
GGUF uses llama-server. Safetensors use vLLM. The built-in echo model only tests the protocol.
YSK product · MIT · Local model gateway · v1.0.0
YSK Omni
Local Hugging Face weights behind an OpenAI-compatible HTTP gateway. Keep the SDKs, curl, and Open WebUI you already use — the models stay on your machine.
CLI: ysk-omni (short name ysko), default port 3850.
Local weights · OpenAI-compatible gateway
Free to useMIT · install yourself
OpenAI shapeChat · image · speech · video
:3850Gateway and admin
Node.js 20+Global npm install
The models stay local. The API stays familiar.
YSK Omni is an OpenAI-compatible gateway in front of local Hugging Face models. Text runs on llama-server or vLLM. Image, speech, transcription, and video go to workers you point at.
- ClientOpenAI SDK · curl · Open WebUI
- Gateway :3850Auth · limits · queue · admin
- Textllama-server · vLLM · echo
- Media workersImage · speech · transcription · video
The gateway does not launch an external coding CLI. To wrap the local Grok CLI, see the Grok CLI Gateway.
Pull in Catalog, then Load
Pull a GGUF or Hub snapshot in Admin, then Load it. A loaded GGUF stays up on llama-server for chat.
Image, speech, transcription
Set OMNI_IMAGE_URL, OMNI_TTS_URL, and OMNI_STT_URL. Until then, those routes return 501 engine_unconfigured.
Video
Point OMNI_VIDEO_URL at an OpenAI-shaped worker. Without one, the gateway may return a short H.264 sample, not a generated video.
One key, the paths you already know
Base URL http://127.0.0.1:3850. A key can limit models. An empty allow-list means the whole catalog.
curl -sS http://127.0.0.1:3850/v1/models \
-H "Authorization: Bearer omni_live_…"
curl -sS http://127.0.0.1:3850/v1/chat/completions \
-H "Authorization: Bearer omni_live_…" \
-H "Content-Type: application/json" \
-d '{"model":"echo","messages":[{"role":"user","content":"Hello"}]}'POST /v1/chat/completions— includingstream: truePOST /v1/responses— OpenAI Responses, text subsetPOST /v1/messages— Anthropic Messages, also acceptsx-api-key/v1/images/*,/v1/videos,/v1/audio/speech,/v1/audio/transcriptionsGET /healthandGET /ready
Admin login is not a live key
Open the admin panel with a one-time code. Do not paste omni_live_… into the browser.
Encrypted chats
Conversations are stored with AES-256-GCM in SQLite at ~/.ysk-omni/data/gateway.db. Weights live under models/.
Key policy
Keys can restrict models. Image, video, and speech need an agent or admin key, unless the caller is an admin one-time session.
Limits and queue
Over the limit returns 429. A full queue returns 429; a draining queue returns 503.
A model list in a few commands
Requires Node.js 20 or newer. Full manual: docs/en.
npm install -g ysk-omni
ysk-omni setup
ysk-omni admin otp
ysk-omni start
# http://127.0.0.1:3850/admin/Free to install. Quoted if we do it for you.
This page has no prices. If you want YSK Limited to install Omni, attach media workers, or adapt it to your environment, we quote that scope in writing.
Questions
Direct answer: Yes. It is MIT licensed. YSK Limited does not charge you to install and use it yourself.
Direct answer: The Grok CLI Gateway wraps the local Grok CLI. YSK Omni runs local Hugging Face weights through llama-server or vLLM.
Direct answer: Port 3850. The CLI is ysk-omni, short name ysko. Admin opens with a one-time code.
Direct answer: Image, speech, and transcription return 501. Without a video worker, you may only get a short H.264 sample, not a generated video.