quickstart

name	quickstart
description	Start one NVIDIA Content Agents REST service locally with Docker Compose. Use when the user asks for /quickstart, local POC setup, or single-service Material, Physics, and Texture startup.
version	0.1.0
author	NVIDIA Content Agents
tags	["content-agents","quickstart","docker","local-deploy"]
tools	["Shell","Docker","curl","Filesystem"]
compatibility	Requires Docker daemon, Docker Compose v2.24+, curl, and repo-root .env provider credentials.

Content Agents Quickstart

When to Use

• Use when the user asks for /quickstart or help starting a local Content Agents proof of concept.
• Use when the user wants exactly one Material, Physics, or Texture REST service from this repository's compose files.
• Use when the user needs a safe preflight, startup, health-check, and follow-up-command handoff for a local service.

Scope

This skill starts one local Content Agents REST service from the repo root. It is a thin wrapper over the existing per-agent Docker Compose files; do not create a new compose stack for this workflow. Material Agent is the recommended first POC.

• Start one target at a time: material, physics, or texture.
• Use deploy-collection when the user wants to run Material, Physics, and Texture together or configure shared dependency endpoints.
• Use the service-specific deploy skills for local NIM sidecars, image-gen sidecars, embedding sidecars, multi-GPU layouts, or deliberate port changes.

Limitations

• Keep secrets out of chat. Tell the user to edit .env; never ask them to paste API keys.
• Always use the repo-root .env file and keep --env-file .env in compose commands.
• Stop before startup when .env was just created or is missing. The user must edit provider credentials before any compose command starts services.
• Do not create a new compose stack for this workflow.
• Do not hold the Shell tool in long build or warm-up loops. Start detached, run only a short bounded health pass, and return clear follow-up commands when the service is still building or warming.

Instructions

1. Infer the target from the user request, or ask for material, physics, or texture when unclear.
2. Check prerequisites and create .env from .env_example only when the template exists.
3. Run the preflight commands and stop before startup when Docker, Compose, .env, required GPU/toolkit checks, or running-container conflicts are not resolved.
4. Start only after conflicts are resolved by stopping the owning stack or by restarting the selected target. If the user keeps an overlapping stack running, skip startup for the new target.
5. Start the selected compose stack with the matching prefix from the target table. Use a background handoff for first builds or any shell tool with a short timeout budget.
6. Run one short bounded health pass, including OVRTX for Material or Physics.
7. Return the service URL, docs URL, startup state, health result, and exact follow-up commands from the output format.

Prerequisites

• Docker daemon reachable.
• Docker Compose v2.24+.
• Repo-root .env with the provider key for the selected backend. The default public backend usually needs NVIDIA_API_KEY.
• Material and Physics require an RTX-capable NVIDIA render GPU plus NVIDIA Container Toolkit because their compose files start OVRTX.
• Recommended render-GPU capacity: Material is a 48 GB-class target and Physics is a 16 GB-class target. Smaller RTX GPUs, including 24 GB or 32 GB cards, may still be valid for small-scene POCs; warn the user and ask before proceeding rather than blocking solely on VRAM.
• A100, H100, H200, and V100 are model-serving GPUs, not render-GPU targets for this quickstart. Texture can run CPU-only with hosted backends.

If .env is missing, create it only when the template exists:

if [ -f .env ]; then
  echo "ENV_FILE=present"
elif [ -f .env_example ]; then
  cp .env_example .env
  echo "ENV_FILE=created; edit .env with one provider key before starting"
else
  echo "ENV_FILE=missing; create .env manually with provider keys"
fi

Treat ENV_FILE=created and ENV_FILE=missing as hard stops. Tell the user to edit the repo-root .env, then rerun this quickstart after credentials are in place.

Preflight

Run this from the repo root and parse the KEY=VALUE lines:

if command -v docker >/dev/null 2>&1 && docker info >/dev/null 2>&1; then
  echo "DOCKER=ok"
else
  echo "DOCKER=missing_or_unreachable"
fi

if docker compose version >/dev/null 2>&1; then
  compose_version="$(docker compose version --short 2>/dev/null)"
  compose_version="${compose_version#v}"
  compose_major="${compose_version%%.*}"
  compose_rest="${compose_version#*.}"
  compose_minor="${compose_rest%%.*}"
  if ! [ "$compose_major" -eq "$compose_major" ] 2>/dev/null || ! [ "$compose_minor" -eq "$compose_minor" ] 2>/dev/null; then
    echo "COMPOSE=unparseable (${compose_version}; need 2.24+)"
  elif [ "$compose_major" -gt 2 ] 2>/dev/null || { [ "$compose_major" -eq 2 ] 2>/dev/null && [ "$compose_minor" -ge 24 ] 2>/dev/null; }; then
    echo "COMPOSE=ok (${compose_version})"
  else
    echo "COMPOSE=too_old (${compose_version}; need 2.24+)"
  fi
elif command -v docker-compose >/dev/null 2>&1; then
  echo "COMPOSE=legacy_v1"
else
  echo "COMPOSE=missing"
fi

if nvidia-smi >/dev/null 2>&1; then
  gpu_mem_mib="$(nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits 2>/dev/null | awk 'BEGIN { max = 0 } $1 + 0 > max { max = $1 + 0 } END { if (max > 0) printf "%.0f", max }')"
  gpu_names="$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | paste -sd '|' -)"
  echo "GPU=yes"
  echo "GPU_VRAM_MIB_MAX=${gpu_mem_mib:-unknown}"
  echo "GPU_NAMES=${gpu_names:-unknown}"
  if docker info --format '{{json .Runtimes}}' 2>/dev/null | grep -q '"nvidia"'; then
    echo "CONTAINER_TOOLKIT=ok"
  elif command -v nvidia-container-cli >/dev/null 2>&1 || command -v nvidia-container-runtime >/dev/null 2>&1; then
    echo "CONTAINER_TOOLKIT=installed_not_registered"
  else
    echo "CONTAINER_TOOLKIT=missing"
  fi
else
  echo "GPU=no"
  echo "GPU_VRAM_MIB_MAX=0"
  echo "GPU_NAMES=none"
  echo "CONTAINER_TOOLKIT=skipped"
fi

echo "RUNNING_CONTAINERS=$(docker ps --format '{{.Names}}' 2>/dev/null | grep -E '^(material-agent-service|physics-agent-service|texture-agent-service|ovrtx-rendering-api|physics-ovrtx-rendering-api|ovrtx_rendering_api-ovrtx-rendering-api-1|vlm-nim|llm-nim|image-gen-nim)$' | paste -sd ',' -)"

Stop before startup when Docker or Compose is not ready. For Material or Physics, also stop when there is no NVIDIA GPU, the NVIDIA Container Toolkit is missing or not registered, or GPU_NAMES contains A100, H100, H200, or V100. If VRAM is unknown or below the recommended capacity, warn the user and ask whether to continue with a small-scene POC.

If RUNNING_CONTAINERS is non-empty, ask whether to stop the owning stack or restart the selected target. Ports overlap across the per-agent compose files; do not start a new target while overlapping containers continue to bind those ports. If the user chooses to leave the existing stack running, skip startup and report the existing-stack state.

Target Table

Target	Prefix	Base URL	Health	Docs
Material	docker compose --env-file .env -f apps/material_agent_service/docker-compose.yml	http://localhost:8000	http://localhost:8000/health	http://localhost:8000/docs
Physics	docker compose --env-file .env -f apps/physics_agent_service/docker-compose.yml	http://localhost:8000	http://localhost:8000/health	http://localhost:8000/docs
Texture	docker compose --env-file .env -f apps/texture_agent_service/docker-compose.yml	http://localhost:8001	http://localhost:8001/health	http://localhost:8001/docs

Infer the target from the user request. If unclear, ask:

Which agent should I start: material (recommended), physics, or texture?

Start

Show a concise summary, then start the selected prefix.

Use the foreground form only when the images are already built or the Shell tool timeout is known to be long enough for the first build:

<PREFIX> up -d --build

For first builds or short Shell tool timeouts, hand off the compose build in the background and return immediately. Store transient logs under logs/quickstart because that path is already ignored by git.

STARTUP_ROOT="logs/quickstart/<target>"
STARTUP_DIR="$STARTUP_ROOT/$(date +%Y%m%d-%H%M%S)"
STARTUP_LOG="$STARTUP_DIR/compose.log"
STARTUP_PID="$STARTUP_DIR/compose.pid"
STARTUP_LATEST="$STARTUP_ROOT/latest"
mkdir -p "$STARTUP_DIR" || { echo "STARTUP=blocked; cannot create $STARTUP_DIR"; exit 1; }
printf '%s\n' "$STARTUP_DIR" > "$STARTUP_LATEST" || { echo "STARTUP=blocked; cannot write $STARTUP_LATEST"; exit 1; }
nohup sh -c '<PREFIX> up -d --build' > "$STARTUP_LOG" 2>&1 &
pid="$!"
printf '%s\n' "$pid" > "$STARTUP_PID" || { echo "STARTUP=blocked; cannot write $STARTUP_PID"; exit 1; }
disown "$pid" 2>/dev/null || true
echo "STARTUP=background pid=$pid log=$STARTUP_LOG"

First build can take around 10 minutes. Material and Physics can take several more minutes while OVRTX warms up. Do not wait through that whole period inside one Shell tool call.

Check

Run a short bounded check. If a background build is still running, report STARTUP=building and return the log command instead of blocking.

<PREFIX> ps

STARTUP_ROOT="logs/quickstart/<target>"
STARTUP_LATEST="$STARTUP_ROOT/latest"
STARTUP_DIR="$(cat "$STARTUP_LATEST" 2>/dev/null || true)"
if [ -z "$STARTUP_DIR" ] || [ ! -d "$STARTUP_DIR" ]; then
  STARTUP_DIR="$(find "$STARTUP_ROOT" -mindepth 1 -maxdepth 1 -type d 2>/dev/null | sort | tail -n 1)"
fi
STARTUP_LOG="${STARTUP_DIR:+$STARTUP_DIR/compose.log}"
# STARTUP_PID is the PID-file path; pid below is the running process ID.
STARTUP_PID="${STARTUP_DIR:+$STARTUP_DIR/compose.pid}"

if [ -n "$STARTUP_PID" ] && [ -f "$STARTUP_PID" ]; then
  pid="$(cat "$STARTUP_PID" 2>/dev/null || true)"
  if [ -n "$pid" ] && kill -0 "$pid" 2>/dev/null; then
    echo "STARTUP=building pid=$pid log=$STARTUP_LOG"
    echo "LOG_COMMAND=tail -f $STARTUP_LOG"
    exit 0
  fi
fi

healthy=no
for i in 1 2 3 4 5 6; do
  if curl -fsS --max-time 5 <HEALTH_URL>; then
    healthy=yes
    break
  fi
  sleep 10
done

if [ "$healthy" = yes ]; then
  echo "SERVICE_HEALTH=healthy"
else
  echo "SERVICE_HEALTH=starting_or_unhealthy"
  if [ -s "$STARTUP_LOG" ]; then
    echo "STARTUP_LOG=$STARTUP_LOG"
    tail -n 80 "$STARTUP_LOG"
  fi
  <PREFIX> logs --no-color --tail=80 || true
fi

For Material or Physics, also run a bounded OVRTX check. If it is not healthy yet, report it as warming unless logs show a clear error. OVRTX and Texture both use port 8001; if Texture owns that port, report a port conflict instead of counting http://localhost:8001/health as healthy for OVRTX.

ovrtx=warming
ovrtx_body=""
for i in 1 2 3 4 5 6; do
  ovrtx_body="$(curl -fsS --max-time 5 http://localhost:8001/health 2>/dev/null || true)"
  if printf '%s\n' "$ovrtx_body" | grep -Eq '"gpu_initialized"[[:space:]]*:[[:space:]]*true'; then
    ovrtx=healthy
    break
  fi
  sleep 10
done
echo "OVRTX_HEALTH=$ovrtx"
if [ "$ovrtx" != healthy ] && [ -n "$ovrtx_body" ]; then
  echo "OVRTX_HEALTH_BODY=$ovrtx_body"
fi

Output Format

Report a concise summary with:

• Selected target and compose prefix.
• Service URL and docs URL.
• Startup state: blocked, not_started, background_building, warming, healthy, or unhealthy.
• Health result, including OVRTX health for Material or Physics.
• Any blocker found during preflight, including .env creation or unresolved port conflicts.
• Log file path or log command when startup is still building or warming.

Then return these follow-up commands:

<PREFIX> logs -f
<PREFIX> ps
<PREFIX> down
<PREFIX> down -v

When stopping a stack that was already running, use the compose file that owns the running container. If ownership is ambiguous, inspect:

docker inspect --format '{{ index .Config.Labels "com.docker.compose.project.config_files" }}' <container-name>

Troubleshooting

Symptom	Cause	Fix
DOCKER=missing_or_unreachable	Docker is not installed, not running, or not reachable by the current user.	Stop and ask the user to start Docker or fix daemon access before retrying.
COMPOSE=too_old or COMPOSE=legacy_v1	Docker Compose is older than v2.24 or only legacy docker-compose is installed.	Ask the user to install Docker Compose v2.24+ and rerun preflight.
ENV_FILE=created or ENV_FILE=missing	The repo-root .env still needs provider credentials.	Stop startup, tell the user to edit .env, then rerun this skill.
Material or Physics reports no GPU/toolkit	The selected target needs the local render stack from Prerequisites.	Stop startup and explain that Texture can run CPU-only with hosted backends.
Material or Physics reports low or unknown VRAM	Small-scene POCs may work, but the target is below recommended capacity.	Warn the user and ask before continuing; do not block solely on VRAM.
Service health stays unhealthy after the short check	The service is still building, warming, or a dependency is misconfigured.	Report the current state and show <PREFIX> ps, curl -fsS <HEALTH_URL>, and <PREFIX> logs --no-color --tail=80.
OVRTX is not immediately healthy	OVRTX can warm up after the service process starts.	Report OVRTX_HEALTH=warming and provide the logs command unless logs show a clear failure.
Texture is running while checking Material or Physics OVRTX	Texture and OVRTX both use port 8001.	Report the port owner conflict instead of counting http://localhost:8001/health as OVRTX healthy.
Ports conflict with another stack	The per-agent compose files share ports.	Stop or restart the owning stack before starting the new target; if the user leaves it running, skip startup.