Run OpenClaw with a Local LLM | DGX Spark

Run OpenClaw with a Local LLM

30 MIN

Install a local-first AI agent and connect it to a private OpenAI-compatible model endpoint

Overview Instructions Agent-ready Models Troubleshooting

CAUTION

Before proceeding, review the security risks in the Overview tab. OpenClaw is an AI agent that can access your files, execute commands, and connect to external services. Data exposure and malicious code execution are real risks. Strongly recommended: Run OpenClaw on an isolated system or VM, use dedicated accounts (not your main accounts), and never expose the dashboard to the public internet without authentication.

Step 1
Prepare your environment (Windows / WSL only)

NOTE

Skip this step on Linux hardware platforms.

If you are on Windows, you can install OpenClaw on native Windows or in Windows Subsystem for Linux (WSL). WSL provides a Linux-style environment if your skills need it; native Windows is often simpler and can connect more easily to Windows apps.

If you choose WSL and it is not already installed:

Open PowerShell as Administrator.
Install WSL:

Powershell

wsl --install

Verify:

Powershell

wsl --version

Start WSL:

Powershell

wsl

Step 2
Install OpenClaw

On your hardware platform, open a terminal and run the official install script:

Bash

curl -fsSL https://openclaw.ai/install.sh | bash

On native Windows, follow the install options documented at openclaw.ai if the curl install path does not apply.

After dependencies are downloaded, OpenClaw will show a security warning. Read the risks; if you accept them, use the arrow keys to select Yes and press Enter.

Step 3
Complete the OpenClaw onboarding

Work through the prompts as follows.

Quickstart vs Manual: Choose Quickstart.
Model provider:
- Recommended for a local model: If your backend is not listed yet (or you will configure it after the server is up), go to the bottom of the list and select Skip for now—you’ll configure the model in a later step.
- If your inference backend appears in the list (for example Ollama, LM Studio, or vLLM), you can select it now and fill in API key (leave blank or use a dummy value), Base URL (default is usually fine), and Model ID (must match the exact handle served by the backend).
Filtering models by provider: If prompted, select All Providers. On the next prompt for the default model, choose Keep Current unless you already selected a provider.
Communication channel: You can connect a channel (for example messaging) to use the bot when away from the machine, or select Skip for Now and configure it later.
Skills: We recommend selecting No for now. You can add skills later from the web UI or Clawhub after you’ve tested the basics.
Homebrew: If you are prompted to install Homebrew, select No—Homebrew is for macOS only and is not needed on Linux or Windows for this playbook.
Hooks: We recommend selecting all three for a better experience. Note that this may log data locally; enable only if you’re comfortable with that.
Dashboard URL: The terminal will print a URL for the OpenClaw dashboard. Save this URL (and any access token shown)—you’ll need it to open the web UI.
Finish: Select Yes on the final prompt to complete installation.

You can now open the OpenClaw dashboard in a browser using the URL and token from the installer.

Step 4
Serve a local model

OpenClaw connects to a local, OpenAI-compatible endpoint. Pick a backend that fits your hardware platform, then use the matching subsection. Model recommendations are in the Agent-ready Models tab.

Option A — vLLM (recommended on large-memory Linux hardware platforms)

Use when you want an OpenAI-compatible HTTP server and validated agent-ready recipes.

In a separate terminal, launch the recommended model for your hardware platform from the Agent-ready Models tab (or follow Serve LLMs with vLLM for container setup).
Wait until the server reports startup complete, then verify:

Bash

curl http://localhost:8000/v1/models

You should see your model handle in the returned list. The default OpenClaw baseUrl for this path is http://localhost:8000/v1.

Option B — LM Studio (simple path on discrete-GPU hardware platforms)

Install LM Studio, download a model that fits your VRAM (see the Agent-ready Models table — the 27B example below is for 24GB+ only), and serve it with a large context window (32K minimum; 64K+ recommended when VRAM allows):

Bash

curl -fsSL https://lmstudio.ai/install.sh | bash
lms get qwen/qwen3.6-27b
lms load qwen/qwen3.6-27b --context-length 65536
lms server start

Use the Model ID that matches what LM Studio serves. Prefer this path when you want a GUI-oriented, llama.cpp-backed workflow.

Option C — Ollama

Bash

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen3.6:27b
ollama run qwen3.6:27b

Inside the Ollama session (or via your usual Ollama config), set context high enough for agent use—for example /set parameter num_ctx 65536. Use the exact Ollama model tag as the OpenClaw Model ID.

NOTE

Free as much VRAM as possible before loading the model (close other GPU workloads; enable only the skills you need). Smaller GPUs should use smaller models—see Agent-ready Models.

Step 5
Configure OpenClaw to use the local server

If you already selected a provider during onboarding and chat works, you can skip to Step 6.

Otherwise, open the OpenClaw config file:

Bash

~/.openclaw/openclaw.json

Example with nano:

Bash

nano ~/.openclaw/openclaw.json

Add or update the models section so it includes your provider. Example for a vLLM server (no API key required—any non-empty placeholder works):

JSON

"models": {
  "mode": "merge",
  "providers": {
    "vllm": {
      "baseUrl": "http://localhost:8000/v1",
      "apiKey": "vllm",
      "api": "openai-responses",
      "models": [
        {
          "id": "nvidia/Qwen3.6-35B-A3B-NVFP4",
          "name": "nvidia/Qwen3.6-35B-A3B-NVFP4",
          "reasoning": true,
          "input": ["text"],
          "cost": {
            "input": 0,
            "output": 0,
            "cacheRead": 0,
            "cacheWrite": 0
          },
          "contextWindow": 262144,
          "maxTokens": 8192
        }
      ]
    }
  }
}

Replace id, name, baseUrl, and contextWindow to match the model and backend you launched. The id and name must match the handle served by the backend.

NOTE

If OpenClaw reports an unsupported-endpoint error against the Responses API, change "api": "openai-responses" to the OpenAI chat-completions variant for your OpenClaw version — vLLM always exposes /v1/chat/completions.

If OpenClaw runs in WSL and cannot reach a server on the Windows host, replace 127.0.0.1 / localhost with the Windows host IP as seen from WSL.

If the OpenClaw gateway is already running, restart it so it reloads ~/.openclaw/openclaw.json.

Step 6
Verify the setup

In a browser, open the OpenClaw dashboard URL (and use the access token if required).
Start a new conversation and send a short message.
If you get a reply from the agent, the setup is working.

You can also ask OpenClaw which model it’s using. In the gateway chat UI you can switch models by typing: /model MODEL_NAME.

Step 7
Optional: add skills and learn more

Skills add capabilities but also risk; only enable skills you trust (for example, community-vetted ones). To add a skill:
- Ask OpenClaw to configure a skill, or
- Use the sidebar in the web UI to enable skills, or
- Browse Clawhub for community skills.
For more usage and configuration details, see the OpenClaw documentation.

Resources