Does my team need accounts with anyone?

No. Open WebUI runs multi-user with its own accounts on your network. No provider account, no per-seat licence, no data leaving the building — and when chat-scanning rules order providers to hand over conversations, there is no provider in this chain to order.

Can I use it from my phone?

Yes — Open WebUI works from any browser on your network, and Enchanted serves your box natively on iPhone and Mac. Remote access happens over your own VPN or tunnel, under your control.

Which server do you install — Ollama or llama.cpp?

Ollama by default: it is NVIDIA's official partner on this machine with Spark-specific builds, and it exposes both OpenAI and Anthropic-compatible APIs so nearly everything connects. llama.cpp where you want maximum control — it has its own official NVIDIA playbook.

The front door

ChatGPT, but it's yours

The app your team already knows how to use, pointed at your own model, on your own network. We install Open WebUI with accounts configured — NVIDIA ships a first-party DGX Spark playbook for exactly this pairing — and the alternatives below all work if you prefer a different door.

Chat front-ends

All verified to run fully against a local endpoint. Licence oddities stated, because two of the biggest names have them.

Open WebUI

Fully local

The full ChatGPT experience — chat, RAG, web search, voice, multi-user accounts. NVIDIA ships a first-party DGX Spark playbook for it.

Custom licence: branding stays above 50 users without an enterprise agreement.

Open WebUI License · 148k★

LobeHub

Fully local

An agent-operations console — schedule and report on a team of agents; strong mobile UI.

Community licence · 81k★

AnythingLLM

Mintplex Labs

Fully local

Drop in your documents, get a private assistant that cites them — agents and MCP included.

MIT · 64k★

Jan

Menlo Research

Fully local

ChatGPT that runs 100% offline, with an OpenAI-compatible server built in.

Custom (read before redistributing) · 44k★

LibreChat

Danny Avila

Fully local

The enterprise one: agents, code interpreter, RAG, OAuth2/LDAP multi-user.

MIT · 42k★

Chatbox

Chatbox AI

Fully local

A lightweight desktop client that just works — minimal configuration.

GPL-3.0 · 41k★

LM Studio

Element Labs

Fully local

GUI runtime + model browser — ships a Linux ARM64 build specifically for DGX Spark.

Proprietary (free)

Enchanted

Gluonfield

Fully local

Your own model on your iPhone and Mac, served from the box at home.

Apache-2.0 · 6k★

The servers underneath

What actually serves the model. GB10-specific status stated — two of these need care on this hardware.

Ollama

Fully local

One command to run any model — NVIDIA's official partner on this exact machine, with Spark-specific builds.

MIT · 178k★

llama.cpp

ggml

Fully local

Maximum-control inference with an official NVIDIA DGX Spark playbook.

MIT · 123k★

vLLM

vLLM Project

Fully local

High-throughput concurrent serving — many users off one box.

Containers only on GB10 — stock pip install fails (PyTorch compiles through sm_120).

Apache-2.0 · 88k★

SGLang

Partial

The fastest structured/constrained generation and prefix caching.

GB10 support is a workaround container, not native — tracking issue open.

Apache-2.0 · 31k★

Questions, answered straight

Does my team need accounts with anyone?: No. Open WebUI runs multi-user with its own accounts on your network. No provider account, no per-seat licence, no data leaving the building — and when chat-scanning rules order providers to hand over conversations, there is no provider in this chain to order.
Can I use it from my phone?: Yes — Open WebUI works from any browser on your network, and Enchanted serves your box natively on iPhone and Mac. Remote access happens over your own VPN or tunnel, under your control.
Which server do you install — Ollama or llama.cpp?: Ollama by default: it is NVIDIA's official partner on this machine with Spark-specific builds, and it exposes both OpenAI and Anthropic-compatible APIs so nearly everything connects. llama.cpp where you want maximum control — it has its own official NVIDIA playbook.

Get yours set up