NemoClaw on Jetson AGX Thor: Edge AI Agent Deployment Guide

NemoClaw is NVIDIA’s enterprise security wrapper for agentic AI — the sandboxed runtime that adds kernel-level isolation, YAML policy enforcement, and privacy routing to the open-source OpenClaw agent framework. When NemoClaw runs on a Jetson AGX Thor at the network edge, it becomes an autonomous AI agent that operates without any cloud dependency. The inference engine runs locally on the Jetson’s integrated GPU. The sandbox runs locally. The policy engine runs locally. No tokens, no prompts, and no tool outputs ever leave the device.

Edge deployment is not a scaled-down version of cloud deployment. It is a fundamentally different operational model where there is no cloud dependency and all data stays on the device. The Jetson has no fallback — if the local model cannot handle a task, there is no cloud API to escalate to. The device may run in environments with no network connectivity at all: factory floors, mining operations, military installations, retail locations in rural areas. This article covers the technical path from unboxing a Jetson AGX Thor to running a production NemoClaw agent in a fully air-gapped configuration.

For the hardware comparison that includes Jetson alongside DGX and RTX options, see our NemoClaw Hardware Guide. For the privacy router configuration that enforces local-only inference, see our NemoClaw Privacy Router deep dive. For information about ManageMyClaw and what we do, visit our About page.

Cloud-based AI agents assume reliable, low-latency internet connectivity. This assumption fails in three categories of enterprise environments where NemoClaw edge deployment on Jetson solves a problem no cloud architecture can address.

Manufacturing and industrial. Factory floors in automotive, semiconductor, and heavy manufacturing facilities frequently operate on air-gapped networks for safety and IP protection reasons. SCADA systems, robotic assembly lines, and quality inspection stations are physically isolated from the internet. An AI agent that assists with predictive maintenance, defect classification, or production scheduling must run entirely on local compute.

Retail and field operations. Retail locations in rural areas, mobile field service vehicles, and pop-up operations often have intermittent connectivity. An AI agent that processes inventory, handles customer interactions, or assists field technicians needs to function when the network drops — and it needs to function fully, not in a degraded mode that prompts “please reconnect to continue.”

Defense and classified environments. Military and intelligence operations require systems that cannot, by regulation, connect to public cloud infrastructure. ITAR-controlled environments prohibit data from crossing certain network boundaries. These are not optional security preferences — they are legal requirements with criminal penalties for violations.

The Jetson AGX Thor is NVIDIA’s highest-performance edge computing module. It uses a Blackwell-architecture GPU with 128 TOPS of AI compute, paired with an ARM-based CPU complex with up to 20 cores. The module supports up to 128GB of unified memory shared between CPU and GPU, which eliminates the PCIe bottleneck that limits discrete-GPU systems when swapping model weights between CPU and GPU memory.

The Jetson runs JetPack OS, NVIDIA’s Linux-based operating system built on Ubuntu. NemoClaw’s OpenShell sandbox requires Linux kernel 5.13+ for Landlock support — JetPack 6.x includes kernel 5.15, which satisfies this requirement. The installation process differs from x86 NemoClaw in two ways: the ARM64 binary must be used, and the model quantization must be optimized for unified memory architecture.

Step 1: Transfer NemoClaw and Model Weights (Pre-Air-Gap)

# On the workstation: create an offline installation package
$ nemoclaw package create \
    --arch arm64 \
    --model nvidia/nemotron-3-super-12b \
    --quantization int4-awq \
    --include-dependencies \
    --output /mnt/usb/nemoclaw-jetson-package.tar.gz

# Transfer the package to the Jetson via USB
$ scp /mnt/usb/nemoclaw-jetson-package.tar.gz jetson@192.168.55.1:/opt/nemoclaw/

# On the Jetson: extract and install offline
$ cd /opt/nemoclaw
$ tar -xzf nemoclaw-jetson-package.tar.gz
$ sudo ./install.sh --offline --arch arm64

Step 2: Configure for Air-Gapped Operation

providers:
  default: vllm-local
  vllm-local:
    endpoint: "http://127.0.0.1:8000/v1"
    model: "nvidia/nemotron-3-super-12b"
    max_tokens: 4096
    timeout_seconds: 180

privacy_router:
  mode: air-gapped
  block_all_external: true
  allowed_endpoints: []
  log_blocked_attempts: true

sandbox:
  mode: deny-by-default
  network:
    allow_loopback: true
    deny_all_external: true

telemetry:
  enabled: false
  phone_home: false

Step 3: Verify the Air-Gapped Stack

# Disconnect the network cable (or disable Wi-Fi)
$ sudo nmcli networking off

# Start vLLM with local model
$ vllm serve /opt/nemoclaw/models/nemotron-3-super-12b \
    --quantization int4-awq \
    --max-model-len 32768 \
    --gpu-memory-utilization 0.80 \
    --port 8000 \
    --host 127.0.0.1

# Start NemoClaw
$ nemoclaw start --config /opt/nemoclaw/nemoclaw-config.yaml

# Verify completely offline operation
$ nemoclaw status
Agent PID: 3412
Sandbox: ACTIVE (deny-by-default)
Provider: vllm-local (127.0.0.1:8000)
Privacy Router: air-gapped (all external blocked)
Network: OFFLINE (loopback only)
Status: RUNNING

# Test agent functionality without network
$ nemoclaw run --prompt "Analyze the sensor data in /data/sensors/latest.csv"
[sandbox] Reading /data/sensors/latest.csv (2.1MB)
[inference] Local vLLM response (3.2s, 847 tokens)
[result] Analysis complete. Report saved to /data/reports/analysis-2026-03-20.txt

Manufacturing: Predictive Maintenance Agent

A NemoClaw agent on Jetson AGX Thor monitors vibration sensors, temperature readings, and motor current draw from CNC machines on a factory floor. The agent runs a predictive maintenance model locally, identifying patterns that precede equipment failures. When the agent detects an anomaly, it generates a maintenance work order and pushes it to the local MES (Manufacturing Execution System) through a sandboxed API call to the factory’s internal network. No data leaves the facility.

agent: "maintenance-predictor"
sandbox:
  mode: deny-by-default
  filesystem:
    read:
      - "/data/sensors/"
      - "/data/equipment-specs/"
    write:
      - "/data/reports/"
      - "/data/work-orders/"
  network:
    allowed_endpoints:
      - "192.168.1.50:8080"  # Local MES system
    deny_all_other: true
  tools:
    allowed:
      - "csv_reader"
      - "anomaly_detector"
      - "work_order_generator"
      - "mes_api_client"

Retail: Inventory and Customer Assist Agent

A Jetson-mounted NemoClaw agent in a retail location processes natural language queries from staff about inventory levels, product specifications, and pricing. The agent reads from a local SQLite database synced nightly when the store’s network is available. During business hours, the agent operates fully offline. Staff ask questions through a local terminal or tablet; the agent responds with inventory counts, product comparisons, and reorder recommendations without any cloud API calls.

Industrial: Safety Compliance Agent

An industrial safety agent on Jetson processes camera feeds and sensor data to verify compliance with safety protocols — PPE detection, restricted zone monitoring, gas level alerts. The agent runs computer vision models locally alongside NemoClaw’s reasoning agent. When a violation is detected, the agent generates an incident report, triggers local alarms, and logs the event to the facility’s safety database. The entire pipeline runs without internet connectivity.

Air-gapped deployments eliminate cloud dependency, but they also eliminate cloud convenience. Every operational task that cloud deployments handle automatically — software updates, model updates, log aggregation, monitoring alerts — must be handled through an offline process on edge devices.

Software updates. NemoClaw and vLLM updates must be packaged on a connected workstation, transferred to a USB drive, and installed on the Jetson during a scheduled maintenance window. There is no over-the-air update mechanism for air-gapped Jetsons. Build a quarterly update cadence and version-lock between NemoClaw, vLLM, and the Nemotron model to avoid compatibility issues.

Model refresh. When NVIDIA releases a new Nemotron model version, the update process is the same: download on a connected machine, transfer via USB, test on a staging Jetson, then deploy to production edge devices. Never update the model on a production edge device without testing first — a model that works on x86 may behave differently on ARM64 with different quantization settings.

Log collection. NemoClaw’s audit logs accumulate on the Jetson’s local storage. For compliance, these logs must be periodically collected — either through a scheduled network sync during maintenance windows, or through physical USB retrieval. Set log rotation to prevent storage exhaustion: the Jetson’s onboard storage is limited compared to server-class hardware.

# Configure NemoClaw log rotation on Jetson
$ nemoclaw config set logging.max_size_mb 500
$ nemoclaw config set logging.max_files 10
$ nemoclaw config set logging.rotation daily

# Export logs to USB for compliance collection
$ nemoclaw logs export \
    --from 2026-03-01 \
    --to 2026-03-20 \
    --output /mnt/usb/nemoclaw-logs-march-2026.tar.gz

# Verify log integrity
$ nemoclaw logs verify --file /mnt/usb/nemoclaw-logs-march-2026.tar.gz
Log integrity: VALID (SHA-256 checksums match)
Entries: 47,231
Date range: 2026-03-01 to 2026-03-20

The Jetson AGX Thor is a powerful edge device, but it is not a DGX. The 12B Nemotron model runs well within the Jetson’s compute and memory envelope. The 120B model does not fit. This means edge agents are limited to the capabilities of the smaller model — which is still highly capable for structured tasks but does not match the 120B model’s performance on complex multi-step reasoning, long-form generation, or nuanced language understanding.

Multi-agent orchestration is possible but constrained. Running two or three agents concurrently with the 12B model will consume most of the Jetson’s unified memory. Five or more concurrent agents will require aggressive quantization (INT4 or lower) and reduced context windows, which degrades output quality. If your edge use case requires multi-agent orchestration at scale, the DGX Station in an air-gapped configuration is the appropriate hardware.

There is no fallback. If the local model cannot solve a problem, the agent cannot escalate to a larger cloud-based model. Design edge agent workflows for tasks where the 12B model consistently succeeds. Use the connected workstation during deployment planning to test every expected workflow against the 12B model before deploying to the air-gapped Jetson.