Deploy the material-agent-service locally using Docker Compose with OVRTX rendering. Use when user wants to run material agent with docker, docker compose, set up local deployment, run the service locally with GPU rendering, start material agent containers, or configure VLM provider for docker deployment. Trigger phrases include "docker compose", "docker deploy", "run locally with docker", "start material agent docker", "docker up", "local deployment", "compose up".
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
The command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
File Explorer
2 files
Showing SKILL.md
SKILL.md
Source instructions · Read-only preview
name
deploy-material-agent-docker
description
Deploy the material-agent-service locally using Docker Compose with OVRTX rendering. Use when user wants to run material agent with docker, docker compose, set up local deployment, run the service locally with GPU rendering, start material agent containers, or configure VLM provider for docker deployment. Trigger phrases include "docker compose", "docker deploy", "run locally with docker", "start material agent docker", "docker up", "local deployment", "compose up".
Requires Docker daemon, Docker Compose v2.24+, NVIDIA Container Toolkit, an RTX-class GPU with 48GB+ VRAM for local OVRTX rendering, repo-root .env provider credentials, and free host ports 8000/8001.
Deploy Material Agent Service with Docker Compose
Deploy the material-agent-service and OVRTX rendering API locally using Docker Compose. No NGC login or Kubernetes required for the default setup.
When to Use
Use when the user wants to run material-agent-service locally with Docker Compose.
Use when the user needs the bundled OVRTX rendering API sidecar on a GPU host.
Use when the user wants to configure cloud VLM providers or an optional local VLM NIM sidecar for Material Agent.
Use quickstart for a shorter first local POC, and use deploy-collection when running multiple Content Agents together with shared dependency endpoints.
Limitations
The default stack owns host ports 8000 and 8001. Stop overlapping Material, Physics, Texture, or standalone OVRTX stacks before startup.
Local OVRTX rendering needs an RTX-class GPU with about 48GB+ VRAM; A100/H100/H200/V100 are not supported for this path.
First build and first render are long-running operations. Start Compose normally, then return logs and health commands instead of holding an agent session open indefinitely.
Keep secrets out of chat and commits. Tell the user to edit .env; do not ask them to paste keys.
Prerequisites
Check before deploying:
Docker Compose v2.24+: docker compose version -- required for env_file: required: false long-form syntax
RTX-capable NVIDIA GPU with 48GB+ VRAM for local OVRTX rendering
(L40, L40S, RTX PRO 6000, etc.; not A100/H100/H200/V100): nvidia-smi
NVIDIA Container Toolkit installed: docker run --rm --gpus all nvidia/cuda:12.0-base nvidia-smi
VLM provider API key (at least one): OpenAI, Anthropic, Gemini, or NVIDIA
Image generation key (optional): public default is Gemini, so set GOOGLE_API_KEY if the user plans to use generated reference images or generated material-library textures. Otherwise set MA_IMAGE_GEN_BACKEND to another configured provider.
Instructions
Confirm Docker, Compose, GPU, NVIDIA Container Toolkit, and port availability before starting the stack.
Create or update the repo-root .env with exactly the provider credentials the selected backend needs.
Start the Material Agent compose stack from the repo root.
Wait for both the main service and OVRTX readiness checks before reporting the service ready.
Return service URLs, health state, log commands, and stop commands using the output format below.
Set VLM API Key
Create .env at the repo root (the compose file reads it via env_file: ../../.env):
To run Cosmos Reason2 8B locally instead of using a cloud VLM API:
# NGC login required for VLM NIM imageprintf'%s'"$NGC_API_KEY" | docker login nvcr.io \
--username '$oauthtoken' --password-stdin
# Add NGC_API_KEY to .env for model weight downloadecho'NGC_API_KEY=...' >> .env# Start with vlm profile + multi-gpu overlay# The overlay pins ovrtx-rendering-api to GPU 0 and vlm-nim to GPU 1.# It is REQUIRED here: NIM's profile selector reports "Free GPUs: <None>"# and refuses to launch when ovrtx is also holding a GPU, even with tens# of GB free. See docker-compose.multi-gpu.yml for the pinning details.
docker compose \
-f apps/material_agent_service/docker-compose.yml \
-f apps/material_agent_service/docker-compose.multi-gpu.yml \
--profile vlm up --build
Requires 2 GPUs (48GB each). VLM NIM takes ~15 minutes to start (model compilation).
The multi-gpu overlay also routes MA_VLM_* and MA_LLM_* to the local
vlm-nim sidecar (nvidia/cosmos-reason2-8b), so no extra .env edits
are needed for --profile vlm.
Operations
View Logs
# All services
docker compose -f apps/material_agent_service/docker-compose.yml logs -f
# Specific service
docker logs material-agent-service
docker logs ovrtx-rendering-api
Stop
# Stop all services
docker compose -f apps/material_agent_service/docker-compose.yml down
# Stop and remove session data
docker compose -f apps/material_agent_service/docker-compose.yml down -v
Rebuild After Code Changes
docker compose -f apps/material_agent_service/docker-compose.yml up --build
# Force full rebuild (no cache)
docker compose -f apps/material_agent_service/docker-compose.yml build --no-cache
docker compose -f apps/material_agent_service/docker-compose.yml up
Check Health
curl http://localhost:8000/health # main service
python - <<'PY'
import json
from urllib.request import urlopen
try:
with urlopen("http://localhost:8001/health", timeout=10) as response:
health = json.load(response)
except Exception as exc:
print(f"rendering API unreachable: {exc}")
raise SystemExit(1)
print(json.dumps(health))
if health.get("status") == "unhealthy":
print("rendering API unhealthy")
raise SystemExit(1)
if health.get("gpu_initialized") is True:
print("rendering API ready")
else:
print("rendering API warming")
PY
Resource Requirements
Configuration
GPUs
Compose CPU limit
Compose memory limit
Default (main service + OVRTX)
1
8
16G
+ VLM NIM profile
2
8
16G
The CPU and memory limits above are the explicit material-agent-service limits in apps/material_agent_service/docker-compose.yml. ovrtx-rendering-api and vlm-nim reserve GPUs but do not define CPU or memory limits in the compose files, so size the host with additional headroom for those sidecars.
GPU Configuration
To assign more GPUs to the rendering API, edit docker-compose.yml:
ovrtx-rendering-api:deploy:resources:reservations:devices:-driver:nvidiacount:2# number of GPUscapabilities: [gpu]
Or assign specific GPU IDs:
device_ids: ['0', '1']
Environment Variables
All configurable via .env file at the repo root:
Variable
Default
Description
OPENAI_API_KEY
OpenAI VLM provider
ANTHROPIC_API_KEY
Anthropic VLM provider
GOOGLE_API_KEY
Google Gemini VLM provider
NVIDIA_API_KEY
NVIDIA (build.nvidia.com) VLM provider
NGC_API_KEY
NGC auth (only for VLM NIM profile)
MA_IMAGE_GEN_BACKEND
gemini
Shared backend for generated reference images and generated material-library textures (gemini, openai, nim)
MA_IMAGE_GEN_MODEL
Optional shared image-generation model override
MA_IMAGE_GEN_BASE_URL
Optional shared image-gen API base URL override
MA_MAX_ACTIVE_SESSIONS
1
Max concurrent pipelines for local OVRTX compose
MA_MAX_RENDER_NUM_WORKERS
1
Max accepted render worker override for local OVRTX compose
WU_NVCF_GLOBAL_MAX_CONCURRENT_REQUESTS
1
Process-wide outbound render request cap
MA_SESSION_TTL_HOURS
24
Session expiry time
MA_MAX_UPLOAD_SIZE_MB
500
Max USD upload size
OVRTX_NUM_SENSOR_UPDATES
500
Sensor update count before capture
NIM_MAX_MODEL_LEN
131072
VLM NIM context length (for 48GB GPUs)
Prediction concurrency is set per request with vlm_max_workers; it is not a
service-wide environment variable.
Generated reference images and generated material-library mode are optional. If
enabled from the client or UI, the service uses MA_IMAGE_GEN_BACKEND=gemini
by default.
Output Format
When handing control back to the user, report:
SERVICE_URL: http://localhost:8000
DOCS_URL: http://localhost:8000/docs
SERVICE_HEALTH: healthy, starting, or unhealthy
OVRTX_HEALTH: healthy only when /health contains "gpu_initialized":true; otherwise warming or unhealthy
If shader compilation messages appear, wait ~5 minutes (first-time only).
Main service unhealthy before rendering API ready
The main service depends_on the rendering API health check. If rendering takes long to start (shader compilation), increase start_period in docker-compose.yml.
VLM NIM KV cache error on 48GB GPUs
Already configured with NIM_MAX_MODEL_LEN=131072 by default. To reduce further:
