Local-first developer dashboard for the NVIDIA DGX Spark.

This project is built for DGX Spark hardware. If you don't own a DGX Spark (GB10 Grace Blackwell, 128 GB unified memory), this repo is not useful to you. The tools assume Spark-local model storage, nvidia-smi access, and memory budgets specific to the GB10. There is no hosted version.

Eight tools that run on your Spark and give you a single UI for the things you'd otherwise do across a dozen terminal windows:

On the same LAN, hit http://<spark-ip> directly:

graph LR
    Browser[Browser] -->|"http://spark-ip"| Nginx[nginx :80]

    subgraph spark [DGX Spark]
        Nginx -->|/| Next[Next.js :3000]
        Nginx -->|/api/| FastAPI[FastAPI :8000]
        FastAPI --> HF[HuggingFace Hub]
        FastAPI --> GPU[nvidia-smi / psutil]
        FastAPI --> NeMo[nemo-automodel / nemo-curator]
        FastAPI --> DD[data-designer]
    end

Off-network, connect via NVIDIA Sync (macOS app) or Tailscale:

graph LR
    Mac[DGX Spark Manager] -->|"NVIDIA Sync"| Spark
    Browser[Browser] -->|"https://spark.tailnet.ts.net"| TS[tailscale serve]
    SSH[Terminal] -->|"tailscale ssh spark"| TSSH[tailscale ssh]

    subgraph Spark [DGX Spark]
        TS --> Nginx[nginx :80]
        TSSH --> Shell[Shell]
        Nginx -->|/| Next[Next.js :3000]
        Nginx -->|/api/| FastAPI[FastAPI :8000]
    end

DGX Lab is self-hosted and local-only -- a Mac browser talking to a DGX Spark on the same LAN or over a Tailscale tunnel. There is no public deployment. CORS is set to allow_origins=["*"] because the only clients are your own devices on a private network.

• NVIDIA DGX Spark
• Python 3.12+ with uv
• Bun 1.3+

SSH into your DGX Spark (or open a terminal on it directly) and clone:

git clone https://github.com/jxtngx/dgx-lab.git ~/dgx-lab
cd ~/dgx-lab

cd backend && uv sync && cd ..
cd frontend && bun install && cd ..
make dev

Open http://localhost:3000 on the Spark, or http://<spark-ip>:3000 from your Mac on the same LAN.

make build
make up

nginx serves on port 80. Stop with make down, rebuild with make rebuild, logs with make logs.

DGX Lab reads and writes to these paths on the Spark (all overridable via env vars):

Hardware constants default to Spark specs and can be overridden:

On the same LAN, open http://<spark-ip> in any browser.

Off-network, two options:

dgx-lab/
├── frontend/
│   ├── apps/web/          # Next.js app (routes, components, hooks)
│   ├── packages/ui/       # Shared UI components (shadcn, Tailwind)
│   └── package.json       # Turborepo workspaces
├── backend/
│   ├── app/main.py        # FastAPI app, CORS, router registration
│   ├── app/config.py      # DGX_LAB_* env vars, memory constants
│   ├── app/routers/       # One router per tool (8 total)
│   ├── app/mock_data/     # Mock data for development
│   └── pyproject.toml     # uv-managed dependencies
├── docker-compose.yaml    # frontend + backend + nginx
├── nginx.conf             # Reverse proxy config
├── Makefile               # dev, up, down, build, logs, rebuild
└── docs/                  # Setup and usage guides

DGX Lab is built and maintained with Cursor and Anthropic Claude as core development agents. The .cursor/agents/ directory contains 14 specialized agent definitions (Chief Architect, Scrum Master, Designer, Backend Engineer, AI Engineer, ML Engineer, Agents Engineer, GOFAI Engineer, AWS Engineer, Tailscale Engineer, Technical Writer, Developer Advocate, DGX Spark Expert, macOS Expert) that encode project conventions, subsystem boundaries, and the DGX Lab design system. They are the persistent context that keeps contributions aligned without a style guide someone has to remember to read. The Developer Advocate agent is there to help anyone who clones or forks the repo get oriented, troubleshoot setup, and understand the architecture well enough to work independently.

• NVIDIA DGX Spark — build.nvidia.com
• DGX Spark User Guide
• NVIDIA NeMo Framework
• LangChain Documentation
• LangGraph Documentation
• LangSmith Documentation

Apache 2.0