--- title: "Run OpenClaw with a Local LLM" publisher: "nvidia" type: "playbook" updated: "2026-08-06T18:38:15.282Z" description: "Install a local-first AI agent and connect it to a private OpenAI-compatible model endpoint" canonical: "https://build.nvidia.com/playbooks/openclaw.md" --- # Basic idea OpenClaw is a **local-first** AI agent that runs on your machine. It remembers conversations, adapts to your usage, runs continuously, uses context from your files and apps, and can be extended with community **skills**. Running OpenClaw with a **local LLM** keeps your data private and avoids ongoing cloud API costs. Your hardware platform provides the GPU acceleration that agent workflows need for responsive tool calling and multi-turn sessions. # What you'll accomplish You will install OpenClaw on your hardware platform, connect it to a local OpenAI-compatible model endpoint, and verify the agent in the OpenClaw web UI. Optionally, you can add communication channels and skills. The agent and models run on your hardware—no data leaves your machine unless you add cloud or external integrations. # Popular use cases - **Personal secretary**: With access to your inbox, calendar, and files, OpenClaw can help manage your schedule, draft replies, send reminders, and find meeting slots. - **Proactive project management**: Check project status over email or messaging, send status updates, and follow up or send reminders. - **Research agent**: Combine web search and your local files to produce reports with personalized context. - **Install helper**: Search for apps/libraries, run installations, and debug errors using terminal access (larger models recommended). # What to know before starting **Required:** - Basic use of the terminal and a text editor - Awareness of the security considerations below **Optional:** - Familiarity with a local inference backend (vLLM, Ollama, or LM Studio) if you plan to use a local model # Important: security and risks AI agents can introduce real risks. Read OpenClaw’s guidance: [OpenClaw Gateway Security](https://docs.openclaw.ai/gateway/security). Main risks: 1. **Data exposure**: Personal information or files may be leaked or stolen. 2. **Malicious code**: The agent or connected tools may expose you to malware or attacks. You cannot eliminate all risk; proceed at your own risk. **Critical security measures:** - **STRONGLY RECOMMENDED:** Run OpenClaw on a dedicated or isolated system (for example, a clean hardware platform or VM) and only copy in the data the agent needs. Do not run this on your primary workstation with sensitive data. - Use **dedicated accounts** for the agent instead of your main accounts; grant only the minimum access it needs. - Enable only **skills you trust**, preferably those vetted by the community. Skills that provide terminal or file system access increase risk significantly. - **CRITICAL:** Ensure the OpenClaw web UI and any messaging channels are **never exposed** to the public internet without strong authentication. Use SSH tunneling or VPN if accessing remotely. - Where possible, **limit internet access** for the agent using firewall rules or network isolation. - **Monitor activity**: Regularly review logs and commands executed by the agent. # Supported hardware platforms Use the matrix below to confirm your hardware platform, OS, memory, and whether multi-node applies. | Hardware platform | OS | Memory | Multi-node capable hardware | | :---- | :---- | :---- | :---- | | **DGX Spark** | DGX OS (Linux) | 128 GB Unified Memory | — | | **RTX** | Linux, Windows, WSL | Dedicated VRAM (size varies) | — | > [!NOTE] > Only platforms listed in the Supported hardware platforms table above are covered by this playbook. # Prerequisites **Hardware requirements** - Supported hardware platform — see Supported hardware platforms matrix above - Enough memory for your chosen model (see **Agent-ready Models** for recommendations) **Software requirements** - Terminal access to the hardware platform (local or SSH) - A local OpenAI-compatible inference backend, or willingness to install one during the Instructions tab - Network access to download OpenClaw and your chosen model checkpoint # Find model guidance This playbook is an **agent** workflow, not a general inference-serving guide. For recommended local models by hardware platform, use the **Agent-ready Models** tab. For full vLLM container recipes and launch settings, see [Serve LLMs with vLLM](https://build.nvidia.com/playbooks/vllm). For Ollama or LM Studio setup, see those playbooks in **Resources**. # Time & risk - **Estimated time:** 30 MIN for install and first-time model setup; model download time depends on size and network - **Risk level:** **Medium to High**—the agent has access to whatever files, tools, and channels you configure. Risk increases significantly if you enable terminal/command execution skills or connect external accounts. Without proper isolation, this setup could expose sensitive data or allow code execution. **Always follow the security measures above.** - **Rollback:** Stop the OpenClaw gateway and uninstall via the same install script or by removing its directory; stop the local inference server separately if desired - **Last Updated:** 07/27/2026 - Added supported hardware matrix and Agent-ready Models tab for DGX Spark and RTX ## More - [Instructions](/playbooks/openclaw/instructions.md) - [Agent-ready Models](/playbooks/openclaw/agent-ready-models.md) - [Troubleshooting](/playbooks/openclaw/troubleshooting.md)