OpenClaw Setup: Run Your Own AI Agent Assistant on Local Hardware (2026)

OpenClaw is a self-hosted gateway that turns your own machine into an always-on AI assistant you can message from the apps you already use — Telegram, WhatsApp, Discord, Signal, Slack, and more. It is open source (MIT), has no paid tier, and runs entirely on hardware you control. The setup takes about five minutes to install and a little longer to make it yours. This guide walks the full path: install, onboarding, the Gateway service, your first channel, and — the part that matters most for this site — pointing OpenClaw at a local model so nothing you say goes to a cloud API.

What is OpenClaw?

OpenClaw runs a single Gateway process on your machine. That Gateway is the bridge between your chat apps and an AI agent: it handles sessions, routing, and channel connections, and it exposes a web dashboard (the Control UI) so you can also chat from a browser. One Gateway serves every channel you configure at once.

Requirements

• Node.js 24.16+ (Node 26 is recommended). The installer can install Node for you if it is missing.
• A model provider — either an API key from a hosted provider, or a local Ollama server running on the same machine (recommended for privacy).
• ~5 minutes, plus whatever hardware your chosen model needs. If you are using a local model, that is the same VRAM math as the Ollama guide.

Install OpenClaw

macOS / Linux / WSL2

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

Windows (PowerShell)

iwr -useb https://openclaw.ai/install.ps1 | iex

The installer detects your OS, installs Node if needed, installs OpenClaw, and then starts onboarding automatically. On Windows, WSL2 is the smoothest path if you also want a Linux toolchain; the native PowerShell install works too. Other install methods (npm, pnpm, bun, Docker, Nix, from source) are in the official docs.

Onboarding

The installer drops you into onboarding, which offers Quick start and Custom setup. Quick start reuses any AI access it detects, verifies it with a real completion, and opens the web dashboard with the Gateway running in the foreground. Custom walks the full guided flow. If you want the classic step-by-step wizard instead:

openclaw onboard --classic

Complete onboarding before anything else — it is what writes your base config and confirms the model actually responds.

Run the Gateway as a service

Quick start leaves the Gateway in the foreground so you can see its logs. Once it is doing what you want, stop it (Ctrl+C) and install it as a background service so it survives reboots:

openclaw gateway install

Open the dashboard

The Control UI is a browser dashboard for chat, config, and sessions. Start it with:

openclaw dashboard

By default it is served locally at http://127.0.0.1:18789/. For remote access use the documented web surfaces or Tailscale — do not port-forward 18789 to the internet (see security below).

Connect your first channel

Telegram is the fastest to wire up and is the best first channel: create a bot with BotFather, add its token in onboarding/config, and you can message your assistant from your phone in minutes. WhatsApp is more involved but very convenient once paired — and it is the one you most want to lock down, because it is tied to your real number.

Supported channels include Discord, Google Chat, iMessage, Matrix, Microsoft Teams, Signal, Slack, Telegram, WhatsApp, Zalo, and more via plugins. You can run several at once from the same Gateway.

Point OpenClaw at your local Ollama model

This is the section that makes it yours. OpenClaw supports local / OpenAI-compatible model services. Since Ollama exposes an OpenAI-compatible endpoint at http://localhost:11434/v1, you point OpenClaw’s provider config there and it uses your local model instead of a cloud API.

1. Make sure Ollama is running and a model is pulled (see the Ollama Quickstart). A 7B–8B model is a fine starting point for an agent.
2. Open your config at ~/.openclaw/openclaw.json.
3. Set the provider’s base URL to http://localhost:11434/v1 and the model name to one you have pulled (e.g. qwen2.5:7b). Ollama does not require an API key, so a placeholder value is enough for the key field.
4. Restart the Gateway and send a test message from the dashboard or your channel.

Security (read this part)

A self-hosted agent that can run tools and answer messages from the internet is a real attack surface if you leave it wide open. Keep these defaults:

• Allowlist who can talk to it. For WhatsApp, set channels.whatsapp.allowFrom to your number(s) only. Example config:

{
  channels: {
    whatsapp: {
      allowFrom: ["+15555550123"],
      groups: { "*": { requireMention: true } }
    }
  },
  messages: { groupChat: { mentionPatterns: ["@openclaw"] } }
}

• Require a mention in groups so it does not react to every message in shared chats.
• Keep it patched. OpenClaw ships security fixes in its releases (recent ones have addressed CVEs such as CVE-2026-25253). Run the latest version and update regularly.
• Use Tailscale for remote access instead of exposing ports 18789 or 11434 to the public internet. A tailnet gives you secure phone-to-home access without opening anything up.

Where to go next: skills & automation

Once the basics work, OpenClaw gets interesting fast. Skills let you teach the agent repeatable procedures it loads on demand. Cron jobs run work on a schedule and deliver results to your chat. There is also a dedicated browser tool for web tasks and multi-agent routing to separate workspaces. These are where a personal assistant starts feeling like infrastructure — explore them in the docs once your first channel is stable.

Troubleshooting

• Installer complains about Node — OpenClaw wants Node 24.16+ (26 recommended). Check with node --version and upgrade if it is older.
• Gateway won’t stay running — run it in the foreground first (openclaw dashboard) and read the error before installing it as a service.
• Model not responding / onboarding fails verification — if you are using a local model, confirm Ollama is up (ollama list) and that the base URL is exactly http://localhost:11434/v1.
• Channel won’t pair (WhatsApp) — pairing is one-time and tied to your number; complete the QR/verification step on the device, then confirm allowFrom includes that number.

FAQ

• Is OpenClaw free? Yes — MIT licensed, no paid tier, developed by the OpenClaw Foundation (a 501(c)(3)). Your only cost is the hardware and, if you use a hosted model, that provider’s API.
• Do I need a powerful machine? Only as much as your model needs. A local 7B–8B model runs on modest hardware; bigger models need the VRAM from the Ollama guide.
• Can I use a cloud model instead of local? Yes — point the provider config at OpenAI, Anthropic, or an aggregator and supply the API key. Local is just the privacy-first option.
• One assistant or many? Multi-agent routing lets you run isolated agents with separate workspaces, sessions, and responsibilities from one Gateway.

Some links above are affiliate links. If you buy through them, ghettoboxai.com may earn a commission at no extra cost to you.