What is Open Claw?

OpenClaw is a self-hosted, always-on AI assistant developed by Peter Steinberger that runs on local environments and private servers. Unlike traditional chat AIs that only respond when called, OpenClaw uses a heartbeat function to check the task list at regular intervals and evaluate items that require action. It is an always-on agent that works autonomously in the background and only notifies humans when a decision is needed.

The rapid adoption of OpenClaw has raised concerns from security researchers regarding model isolation and access control to local data. NVIDIA collaborated with Steinberger and the developer community to address these issues, resulting in the release of NemoClaw. NemoClaw integrates OpenClaw, the NVIDIA OpenShell Secure Runtime, and the NVIDIA Nemotron™ Open Model to help organizations securely deploy Claw.

Reference: Nemotron Labs: Benefits of OpenClaw agents for various organizations

What is the Explog Menu?

Hermes Agent is an agent developed by Nous Research, based on the Hermes model family, and specializes in inference and dialogue. According to OpenRouter, it is the most widely used agent in the world, and has received over 140,000​ ​GitHub stars in just three months.

It features self-evolution capabilities that autonomously learn and save skills from tasks and feedback, and sub-agent capabilities that divide complex tasks into dedicated workers. When combined with large-scale models such as the NVIDIA Nemotron 3 Ultra, it delivers high inference accuracy.

Reference: Hermes develops self-improving AI agent using NVIDIA RTX PC and DGX Spark.

What is NVIDIA Nemotron 3 Ultra?

NVIDIA Nemotron 3 Ultra is the LLM used in this article. It is a MoE hybrid Mamba-Attention architecture with a total of 550 parameters and 550 active parameters, supporting contexts of up to 1 million tokens. It has achieved 91% in the agent productivity metric PinchBench and 95% in the long-context benchmark RULER (1M tokens), achieving up to 5 times the throughput of open models in its class.

For operation with vLLM, we will use the NVFP4 quantized version. The license is OpenMDW-1.1 from the Linux Foundation.

Reference: NVIDIA Nemotron 3 Ultra - NVIDIA Nemotron

　　　NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents | NVIDIA Technical Blog
　　　nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 · Hugging Face

This article explains how to run Nemotron 3 Ultra and NemoHermes together, as well as its use cases. Furthermore, because this verification uses a large parameter size model of 550B, we utilized KDDI Corporation's KDDI GPU Cloud. By leveraging Nemotron 3 Ultra on KDDI GPU Cloud, which boasts high computing power and NVIDIA AI Enterprise, it is possible to promote the implementation of highly accurate AI agents while ensuring sovereignty with a fixed-fee pricing model that facilitates budget planning.

Reference: KDDI GPU Cloud

Verification environment

1-1. Preparing the local cache directory

Set the NGC_API_KEY as an environment variable and create a directory for the model cache. Obtain the NGC_API_KEY from NVIDIA NGC.

export NGC_API_KEY=<PASTE_API_KEY_HERE>
export LOCAL_NIM_CACHE=~/.cache/nim
mkdir -p "$LOCAL_NIM_CACHE"
chmod -R a+w "$LOCAL_NIM_CACHE"
export SERVED_MODEL_NAME=nvidia/nemotron-3-ultra-550b-a55b

1-2. Starting LLM

Start NVIDIA NIM with the following command:

docker run --rm -d \
  --gpus '"device=0,1"' \
  --shm-size=16GB \
  --ipc=host \
  -p 8000:8000 \
  -v "$LOCAL_NIM_CACHE:/opt/nim/.cache" \
  -e NIM_MODEL_PATH=hf://nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 \
  -e NIM_SERVER_PORT=8000 \
  -e NGC_API_KEY="$NGC_API_KEY" \
  -u "$(id -u)" \
nvcr.io/nim/nvidia/model-free-nim:2.0.6 \
nim-serve \
  --served-model-name "$SERVED_MODEL_NAME" \
  --tensor-parallel-size 2 \
  --kv-cache-dtype fp8 \
  --max-model-len 262144 \
  --reasoning-parser nemotron_v3 \
  --enable-auto-tool-choice \
  --mamba-backend flashinfer

1-3. Checking LLM Status

Check if LLM started correctly.

curl -X 'POST' \
  'http://0.0.0.0:8000/v1/chat/completions' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "nvidia/nemotron-3-ultra-550b-a55b",
    "messages": [{
      "role": "user",
      "content": "Which number is larger, 9.11 or 9.8?"
    }],
    "max_tokens": 1024
  }' | jq -r '.choices[0].message.content'

Note: jq is not required. Install it as needed using the instructions below.

sudo apt install -y jq

If a normal response is received, the LLM is ready.

2-1. Creating a Discord Account

Please create a Discord account using the link below.
https://discord.com/

2-2. Creating a Discord Server

Create a new server by clicking the "+" icon in the left sidebar of the Discord screen.

2-3. Add a Bot to the Server

4. Go to OAuth2 → URL Generator.

- Select the bot in Scopes.
- Grant appropriate text permissions in Bot Permissions.
Copy the generated URL and open it in your browser to add the bot to the server.

3-1. Executing the installation command

export NEMOCLAW_AGENT=hermes
export NEMOCLAW_GATEWAY_PORT=8082
export NEMOCLAW_PROVIDER=vllm
export NEMOCLAW_VLLM_PORT=8000
export NEMOCLAW_MODEL=nvidia/nemotron-3-ultra-550b-a55b
export NEMOCLAW_SANDBOX_NAME=hermes
export NEMOCLAW_HERMES_DASHBOARD=1
curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash

3-2. Setting up a messaging channel

Select Discord
Enter the Bot Token you copied earlier.
Right-click the server icon in the sidebar, copy the Server ID, and paste it into the input field.

3-3. Complete the remaining steps.

Follow the on-screen instructions to complete the remaining settings in order.

3-4. Access to NemoHermes

Once the installation is complete, you will see output similar to the following:

----------------------------------------------------
Hermes is ready
Sandbox:  hermes
Model:    nvidia/nemotron-3-ultra-550b-a55b (Local vLLM)
Access
Hermes Agent OpenAI-compatible API
Port 8642 must be forwarded before connecting.
http://127.0.0.1:8642/v1
Hermes Agent Web dashboard
Port 9119 must be forwarded before opening this URL.
http://127.0.0.1:9119/
Terminal:
  nemohermes hermes connect
Manage later
  Status:      nemohermes hermes status
  Logs:        nemohermes hermes logs --follow
  Model:       nemohermes inference set --model <model> --provider <provider> --
  Policies:    nemohermes hermes policy-add
  Credentials: nemohermes credentials reset <KEY> && nemohermes onboard
----------------------------------------------------
[INFO] === Installation complete ===

You can now send messages to NemoHermes via Discord.

NemoHermes boot screen after onboarding is complete.

What is VSS (Video Search and Summarization)?

VSS stands for "NVIDIA Metropolis Blueprint for Video Search and Summarization," a video analytics platform that integrates real-time feature extraction and analysis, incident detection and alert verification, and agent-based processing such as natural language video search, Q&A, and summarization for video and live streams. It is suitable for applications such as automated analysis of surveillance camera footage in warehouses and factories, anomaly detection, and SOP (Standard Operating Procedure) verification. For more details, please refer to NVIDIA VSSBlueprint (GitHub).

Automate video surveillance with VSS x NemoHermes

By integrating NemoHermes with VSS, the following uses are possible:

Incident Confirmation

直近1時間のインシデントの内訳を教えて

Using NemoHermes' built-in cron, you can perform video monitoring at regular intervals and automatically send notifications if any abnormalities are detected.

- The sensor status is checked at a fixed time every hour/day, and a notification is sent to Discord only if an abnormality is detected.
- The system automatically patrols during unmanned hours at night and on holidays, and contacts the person in charge when an alert is triggered.
- Check surveillance cameras at multiple locations sequentially and report any abnormalities.

Generate PDF report

直近1時間のインシデントをレポートにしてください

Furthermore, NemoHermes supports specifying report formats. By providing instructions to match your company's unique report format (daily reports, weekly reports, safety checklists, etc.), you can generate reports that can be submitted directly by simply copying and pasting.

昨日の異常イベントを、以下のフォーマットで日報として出力してください。 ーーーーーーー 【日報】YYYY年MM月DD日 安全巡視結果 1. 巡視エリア： 2. 検出された異常： 3. 対応状況： 4. 今後の注意点： ーーーーーーー

This allows reporting tasks that previously required manual work and took considerable time to be completed in seconds, significantly reducing the workload on field staff.

Natural language cluster management enabled by Kubernetes × NemoHermes

Managing Kubernetes clusters using natural language

By using an MCP server, you can control your Kubernetes cluster using natural language from chat tools such as Discord.
For example, checking the cluster status can be done with just the following instructions:

k8s-mcp のスキルを使ってクラスタの状態について分析を行って

NemoHermes investigates the entire cluster through the Kubernetes MCP Server.

Node status
Pod status
· event
Log
• Overall system health

We will analyze these factors and report the results.

In this example, the storage-provisioner anomaly was automatically detected.
Furthermore,

storage-provisioner の問題を修正して

When instructed to do so, NemoHermes

1. Obtain relevant logs.
2. Analyze the cause
3. Implement appropriate corrective measures.
4. Reconfirm the corrected state.

This series of tasks will be performed automatically.

Previously, troubleshooting and recovery tasks that required specialized Kubernetes knowledge can now be performed using only natural language.

Deploying applications to a Kubernetes cluster

MCP servers can be used not only for troubleshooting but also for application deployment.
For example, you can give the following instructions from Discord:

nginx をクラスタにインストールして

NemoHermes uses Kubernetes MCP Server to select the appropriate deployment method.

- Create a Deployment
- Create a Service
- Check if the Pod is running.
- Publish the Service
The process will be performed automatically up to this point.

After deployment is complete, on the cluster nginx You can also verify that it is functioning correctly and that the Service is publicly available.

By combining NemoHermes with Kubernetes MCP Server, you can manage your Kubernetes cluster using natural language. This significantly reduces the operational burden of Kubernetes, and a major advantage of NemoHermes and MCP is that even users without specialized knowledge can manage the cluster.

Reference URL

• NemoClaw Official Documentation
• NemoHermes Quickstart
• OpenClaw Quickstart
• NemoClaw GitHub
• NVIDIA Nemotron™ 3 Ultra – Research Page
• NVIDIA Nemotron™ 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents（NVIDIADeveloper Blog）
• NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4（Hugging Face）
• Hermes AI Agent: Self-Healing AI on NVIDIA Hardware (NVIDIA Japan Blog)
• What the OpenClaw agent means for all organizations (NVIDIA Japan Blog)
• NVIDIA Blueprint for Video Search and Summarization（VSS）