DGX Station AI Skills and dgx-assist

Basic idea

DGX Station AI Skills teaches your AI coding agent how to operate a DGX Station correctly. It installs four native Agent Skills plus a dependency-free dgx-assist command-line tool into a project you choose, so that asking your agent "serve this model" or "why is this GPU unavailable?" produces answers grounded in your actual hardware instead of recalled generic advice.

The skills route the agent through a fixed workflow: inspect the real Station, search a pinned snapshot of NVIDIA guidance, resolve an exact named model to a qualified recipe, run preflight, ask you to approve, then act and verify. Every mutation requires your confirmation, and dgx-assist only ever stops services it started and labelled as its own.

The same CLI works without an agent. Add --json and it emits a stable versioned envelope for scripts and CI; omit it and you get readable headings, tables, and action previews for terminal use.

What you'll accomplish

You'll install the four DGX Station skills and the dgx-assist CLI into a project, then drive a qualified vLLM inference workload from a plain-language request through preflight, approval, launch, and verification.

You'll also be able to:

• Inspect your Station's software profile and see exactly which actions its release qualifies it for.
• Search pinned NVIDIA Development Guide and Bring-Up Guide content with revision and source-digest provenance on every result.
• Plan MIG layouts and run read-only diagnostics that produce redacted support bundles.

What to know before starting

• Experience with the Linux command line and running shell scripts
• Familiarity with an AI coding agent — Claude Code, Codex, Gemini CLI, or Cursor — and how it loads project-level context
• Basic understanding of Docker containers and GPU device selection
• Familiarity with vLLM or SGLang inference serving (helpful but not required)

Two safety rules matter more than the rest, and the skills enforce them for you: select GPUs by UUID rather than by nvidia-smi index, which is not a CUDA ordinal on this platform; and treat an unrecognized software build as unknown rather than assuming it behaves like a qualified one.

Prerequisites

Hardware Requirements:

• NVIDIA Grace Blackwell GB300 Ultra Superchip System (DGX Station)
• GB300 compute capability 10.3, confirmed by nvidia-smi --query-gpu=compute_cap --format=csv
• At least 20GB available storage space for the qualified model weights and container image

Software Requirements:

• Python 3.11 or newer: python3 --version
• Docker with the NVIDIA Container Toolkit: docker info | grep -i runtime
• One supported AI coding agent installed: claude --version, codex --version, gemini --version, or Cursor
• Network access to download model weights and container images on first run
• No inbound port access is required; inference binds to localhost by default

Ancillary files

All required assets can be found in the DGX Station AI Skills playbook repository.

• assets/install.sh — Thin wrapper that runs the Python installer
• assets/installer.py — Installs, updates, migrates, and uninstalls the skills; writes .dgx-station/install-manifest.json and manages the delimited NVIDIA block in your agent context file
• assets/dgx-assist.pyz — The standalone zipapp CLI providing host inspection, guidance search, recipe resolution, preflight, service lifecycle, MIG planning, and diagnostics
• assets/skills/dgx-station/ — Platform compatibility, coherency, GPU selection, container, CDI, and power guidance
• assets/skills/dgx-station-inference/ — Exact-model recipe resolution, preflight, approved launch, verification, and owned service lifecycle
• assets/skills/dgx-station-mig/ — Driver-discovered MIG inspection, planning, approved apply, and restoration evidence
• assets/skills/dgx-station-diagnose/ — Read-only diagnostics, playbook correlation, redacted bundles, and one allowlisted fix at a time
• assets/AGENTS.md — The managed context block written into your agent's project context file

Time & risk

• Estimated time: 15 minutes to install and verify, plus 20–40 minutes if you run the qualified inference recipe (most of that is the first model and container download)
• Risk level: Low
  • The installer refuses unmanaged file collisions and symlink destinations rather than overwriting anything it does not own
  • First-run model and container downloads require network bandwidth and may fail on a slow or interrupted connection
  • Only one model is qualified to launch in v1; larger bundled candidates are reported as non-runnable by design
  • MIG changes are disruptive and are blocked unless your release profile enables them and you explicitly approve the plan
• Rollback: Run assets/install.sh uninstall --target /path/to/project to remove every manifest-owned file and the delimited context block. The installer preserves any managed file you modified and leaves unrelated context untouched. A user-scope CLI installed with install-cli is removed by deleting ~/.local/bin/dgx-assist. Cached state lives under ~/.cache/dgx-assist/ and ~/.local/state/dgx-assist/ and can be deleted.
• Last Updated: 07/28/2026
  • First Publication