Your first steps with Sparkbot.

1. Install on your DGX Spark

Open a terminal on the Spark, or connect to it over SSH. Use your normal Linux account, then run:

curl -fsSL https://sparkbot.dev/install.sh | bash

The installer creates its own Python environment and prepares the compact CPU helper engine. It does not download model weights or start GPU workloads. Missing host packages may require your administrator password. Initial application downloads come from this website; a GitHub account is not required.

2. Open your dashboard

The installer prints a setup link. Open it to create your dashboard password, then visit your Spark’s address on port 10005, such as http://<spark-address>:10005.

The dashboard listens on all interfaces for headless access. Keep it on a trusted network or use HTTPS. For local password or token entry from another computer, you can use an SSH tunnel:

ssh -L 10005:127.0.0.1:10005 user@spark

Then open http://localhost:10005. If you lose the setup link, run ./manage.py setup-link from your installed application folder.

3. Add your local assistant

Open Settings → SparkBotLLM → Set up local helper. The public Q4 model downloads from Hugging Face, about 1.05 GB. The default helper runs on CPU with a 4K context and one request at a time. It uses loopback port 10004.

Ask it to check Docker support, inspect GPU clocks, or explain a deployment requirement. Review proposed commands before they run. Administrator passwords belong in the dashboard approval dialog, never in chat.

4. Choose a recipe

Open the recipe library and check a profile’s hardware, model download and disk requirements. Select Install & run when ready. Sparkbot verifies its pinned model files and checks that the GPUs are free before starting its own containers.

For two Sparks, first add the worker’s verified SSH connection in Settings and connect the supported interconnect. Missing Docker and NVIDIA container support have reviewed setup steps. Existing models and unrelated containers stay under their own managers.

New TensorFold launchers have been checked in code and with fixtures; full GPU deployments on idle hardware remain unrun. DeepSeek preparation has substantial disk requirements. Check the details before installing.

5. Find your working rhythm

Use Overview for aggregate throughput, prompt processing, cache reuse and hardware. Benchmarks offers decode rows at 1, 2, 4 and 8 concurrent requests and separate prefill sizes. Cost figures estimate the API equivalent of observed local usage.

Guided mode explains steps. Expert mode exposes terminal access and supported launch overrides. The helper’s chat window can be resized or hidden while you work. Settings includes reviewed GitHub update checks, which become publicly available once the application repository is public.

Credits

Sparkbot is an independent project by x00byte. SparkBotLLM is adapted from LiquidAI LFM3; llama.cpp supplies the native CPU engine. Model recipes retain ThinkingCap, myllmbox, Mia’s AI Lab, TensorFold/BertholomusAI, model publisher and contributor notices. NVIDIA DGX Spark is a supported platform.