Unbox. Install an agent. Feed it this site.

A new DGX Spark does not need a hand-copied bash novel. Install Grok Build, Codex, or Claude Code, point the agent at llms.txt (or this page), answer the interview, and let it build clustered TP=2 or isolated solo TP=1 from current launcher commands.

Out-of-box flow

Unbox the Spark (or both Sparks), cable power, mgmt LAN, and the 200 Gb/s DAC if you have two.
First install: a coding agent on the Spark or on a laptop that can SSH in — Grok Build, Codex, Claude Code, or similar.
Give the agent llms.txt and/or llms-full.txt. The Pages generator at the root is optional if you would rather download scripts yourself.
The agent asks the required questions. Only then does it clone eugr/spark-vllm-docker, pull vllm-node, download weights, and wire Open WebUI / n8n / Hermes to :8000/v1.

Paste this into the agent

starter prompt

Read https://cmdlabtech.github.io/dgx-spark-ai-stack/llms-full.txt (fallback: https://raw.githubusercontent.com/cmdlabtech/dgx-spark-ai-stack/main/llms-full.txt).

I just unboxed Nvidia DGX Spark hardware. Interview me with every required question in that playbook. Do not apt, docker, ufw, or launch vLLM until I have answered. Do not use documentation IPs, LiteLLM, --tf5, or vllm-node-tf5. After I confirm the plan, build the layout.

If the agent cannot fetch GitHub Pages, paste llms-full.txt into the chat instead of the URL.

What it will ask

The playbook forbids guessing. Expect questions about node count, clustered vs solo, Linux username, mgmt IPs, DAC IPs, Hugging Face token, model preset, apps (Open WebUI, n8n, Hermes), whether to disable the Ubuntu desktop, and whether the agent is running on the Spark or over SSH.

A — Clustered TP=2

One vLLM across both GPUs. Head on node-a :8000. Worker has no API. Typical recipe: qwen3.5-122b-fp8, window 262144. The head API can see prompts from both UIs.

B — Isolated solo TP=1

Each node owns its GB10. Use this for one Spark, or for two Sparks that must not share a brain. INT4 122B window 32768; 35B FP8 window 65536. Do not pass --solo to a cluster_only recipe.

Files the agent should read

/llms.txt — index (this project's map)
/llms-full.txt — full playbook (interview + commands)
docs/architectures.md — layout details and switch procedure
Setup generator — same procedure as downloadable setup-node-a.sh / setup-node-b.sh

The generator and the playbook target current eugr/spark-vllm-docker: image tag vllm-node, build-and-copy.sh without --tf5, hf-download.sh for weights, no-Ray as the multi-node default. Old copies that still say vllm-node-tf5 or LiteLLM are stale.