Run NemoClaw with a local LLM

NemoClaw is an open-source reference stack from NVIDIA that runs OpenClaw with the NVIDIA OpenShell runtime bundled.

This guide walks you through running NemoClaw on Olares with the Gemma4 26B model app as the backend LLM.

Alpha software

NemoClaw is an early preview release from NVIDIA and is not recommended for production use. For official updates and community feedback, see NVIDIA/NemoClaw.

Learning objectives

In this guide, you will learn how to:

• Install and configure NemoClaw with a local LLM.
• Keep the model loaded for always-on responses.
• Start your first chat with the agent.
• Connect the agent to Discord for remote chat.
• Enable real-time web search.

Prerequisites

• A local model app installed and running on your Olares device.
• Admin privileges to install apps from Market and edit application settings.

Get the model name and endpoint URL

NemoClaw needs the model name and its shared endpoint URL during installation.

1. Open your model app from Launchpad and note the model name shown on the page. In this example, it's gemma4:26b.

   

2. Open Settings, then go to Applications > Gemma4 26B Q4_K_M (Ollama).

3. In Shared entrances, select Gemma4 26B Q4_K_M to view the endpoint URL.

   

4. Note the shared endpoint. For example:

   http://2e53d5230.shared.olares.com

   Why the shared endpoint?

   The URL on the model app's main page is user-specific and routes through the browser. The shared endpoint is reachable from other apps on Olares without sign-in or CORS issues, which is what NemoClaw needs.

Install NemoClaw

1. Open Market and search for "NemoClaw".

   

2. Click Get, then Install.

3. When prompted, set the environment variables:

   • NEMOCLAW_ENDPOINT_URL: Enter or paste the shared endpoint URL, and append /v1, such as http://2e53d5230.shared.olares.com/v1.
   • NEMOCLAW_MODEL: Enter or paste the model name, such as gemma4:26b.

   

   TIP

   You can change these environment variables later in Settings > Applications > NemoClaw > Manage environment variables.

4. Click Confirm and wait for installation to complete.

   Installation takes about 15 minutes, depending on your network. During this time, NemoClaw installs the NVIDIA OpenShell runtime and runs the initial agent onboarding.

   WARNING

   Keep the model app running during installation. The initial onboarding requires the model to be reachable, and the installation won't complete if the model stops or becomes unavailable.

When the installation finishes, two shortcuts appear on Launchpad:

• NemoClaw CLI: The terminal interface for running NemoClaw and OpenClaw commands.
• OpenClaw Web UI: The browser-based dashboard for OpenClaw.

Keep the model loaded (optional)

By default, the local LLM unloads from memory after 5 minutes of inactivity, and the next reply has to wait for the model to reload. For an always-on agent, enable the keep-alive setting on the model app to keep it resident in memory.

1. Open Settings and go to Applications > Gemma4 26B Q4_K_M > Manage environment variables.

2. Find KEEP_ALIVE, click edit_square, set the value to true, and click Confirm.

   

3. Click Apply.

When to leave KEEP_ALIVE unset

Keeping the model loaded consumes VRAM continuously. If you only use the agent occasionally and don't mind the cold-start delay, leave KEEP_ALIVE unset.

Start your first chat

NemoClaw lets you chat with your agent in either the OpenClaw Web UI or the OpenClaw TUI inside the NemoClaw CLI. Because the model and endpoint were configured during installation, you can skip the manual onboarding and go straight to a session.

Integrate with Discord

To chat with your NemoClaw agent remotely, connect it to a Discord bot. You need a Discord account and a server where you have permission to add bots.

Step 1: Create a Discord bot

1. Log in to the Discord Developer Portal with your Discord account.

2. Click New Application.

   

3. Enter a name for the new app, agree to the terms, and click Create.

   

4. From the left sidebar, select Bot.

5. Scroll down to the Privileged Gateway Intents section and enable the following settings:

   • Presence Intent
   • Server Members Intent
   • Message Content Intent

6. Click Save Changes.

7. Scroll up to the Token section, click Reset Token, and copy the generated token. You need this token in Step 3.

   

Step 2: Invite the bot to your server

1. From the left sidebar, select OAuth2 and find the OAuth2 URL Generator section.

   a. In Scopes, select Bot and applications.commands.

   

   b. Scroll down to Bot Permissions and configure them as shown. You can adjust these later.

   

2. Copy the Generated URL at the bottom.

3. Paste the URL into a new browser tab, select your Discord server from Add to server, click Continue, and click Authorize.

   The bot is authorized and added to your server.

   

Step 3: Configure the Discord channel

NemoClaw runs OpenClaw inside a sandboxed runtime, so you must configure the channel from within the runtime shell.

1. Open the NemoClaw CLI app from Launchpad.

2. Connect to the runtime sandbox:

   nemoclaw my-assistant connect

   Wait until the terminal shows the sandbox prompt, such as sandbox@my-assistant:~$.

3. Run the channel configuration wizard:

   openclaw config --section channels

4. Follow the prompts to add Discord:

   Settings	Option
   Where will the Gateway run	Local (this machine)
   Channels	Configure/link
   Select a channel	Discord (Bot API)
   How do you want to provide this Discord bot token?	Enter Discord bot token, and paste the token from Step 1.
   Configure Discord channels access	Yes
   Discord channels access	Open (allow all channels)

5. When finished, when prompted to select a channel, select Finished.

6. When prompted to configure DM (Direct Message) access policies, select Pairing.

Discord channel stuck in startup-not-ready

If the Discord channel shows startup-not-ready in the OpenClaw Web UI, restart the gateway. See Common issues for the steps.

Step 4: Authorize your Discord account

For security, the bot doesn't respond to unauthorized users. You must pair your Discord account with the bot.

1. Open Discord and send a direct message to your bot.

   The bot replies with a pairing code and a command.

   

2. Switch back to the NemoClaw CLI sandbox shell and use the command provided by the bot to approve the pairing. For example:

   openclaw pairing approve discord FY6PAVY8

   When you see the following, it means the Discord is authorized:

   Approved discord sender 1277468602303385654.

3. After approval, you can chat with your agent directly in Discord.

   

Enable web search

By default, the agent answers only from its training data. To let it fetch real-time internet information, you can use a web search provider. The following uses SearXNG as an example, which you can install as a self-hosted instance from Olares Market.

1. Install SearXNG from Olares Market.

2. Open Settings, then go to Applications > SearXNG.

3. In Shared entrances, select SearXNG to view the endpoint URL.

   

4. Copy the shared endpoint. For example:

   http://d1236e020.shared.olares.com

5. Open the NemoClaw CLI app from Launchpad.

6. Connect to the runtime sandbox:

   nemoclaw my-assistant connect

7. Run the web tool configuration wizard:

   openclaw config --section web

8. Configure as follows:

   Settings	Option
   Where will the Gateway run	Local (this machine)
   Enable web_search	Yes
   Search provider	SearXNG
   SearXNG Base URL	Paste the shared SearXNG endpoint from Step 4
   Enable web_fetch (keyless HTTP fetch)	Yes

9. To verify, ask your agent a question that requires real-time information. For example:

   What are today's top tech news headlines?

   The agent should fetch and cite live web results.

Install skills

Skills add capabilities to your agent, such as managing Olares files and apps or integrating with Google Workspace.

1. Open the OpenClaw Web UI from Launchpad.

2. Go to Skills.

3. Search for the skill in ClawHub and click Install.

4. Open the chat page in the OpenClaw Web UI and run /reset to start a new session so the agent picks up the newly installed skill. If you've configured channels such as Discord, also run /reset in each channel conversation.

   TIP

   You can also install skills from the NemoClaw CLI sandbox using openclaw config --section skills.

For walkthroughs of common skills, see:

• Manage Olares with Olares CLI: Let the agent operate files and apps on your Olares device through natural language.
• Integrate with Google Workspace: Connect Gmail, Calendar, and Drive via the gog skill.

For more on managing skills, see Manage skills and plugins.

Install plugins

Plugins extend OpenClaw with additional channels and integrations.

1. Open the NemoClaw CLI app from Launchpad.

2. Connect to the runtime sandbox:

   nemoclaw my-assistant connect

3. Install the BlueBubbles plugin:

   openclaw plugins install @openclaw/bluebubbles

For other plugins, use the standard openclaw plugins list and openclaw plugins install <name> commands inside the runtime. For details, see Manage skills and plugins.

Common issues

For a list of common issues and workarounds, see Common issues.

Learn more

• NVIDIA NemoClaw: Official reference stack and documentation from NVIDIA.
• OpenClaw: Set up OpenClaw features such as persona setup.