OpenClaw is an agent host that connects models with tools such as a browser and messaging channels. Its usefulness depends on the model, configuration, and permissions. API access does not remove legal obligations or every provider safeguard.
You need a model-provider account and API key. The original article discusses SiliconFlow (Chinese original); other supported providers are also possible. Check usage limits and enable spending controls where available. Multi-step agents can consume much more context than a short chat.
Choose a Windows installation path
The original February 2026 walkthrough used native PowerShell and Node.js 22 or later. Current Windows documentation also describes Windows Hub and WSL2 options. Follow the requirements for the method you choose; WSL2 provides a Linux-compatible runtime.
The original downloads were Node.js 22.13.1 and Git for Windows 2.47.1. These are historical versions. Prefer a supported release from the official project for a new installation.
The original PowerShell installer
This command downloads and executes OpenClaw’s installer. Review the script before running it. Administrator rights may be needed for some installation steps, but should not be assumed necessary for every later agent operation:
iwr -useb https://openclaw.ai/install.ps1 | iexWhen the installer reports success, follow onboarding to select your provider and model, enter the API key, and configure channels. Use the provider’s exact model identifier. You can revise settings afterward.
Connect Telegram
Open @BotFather, send /newbot, and choose a name and available username. Enter the returned token into the Telegram channel configuration. Keep it separate from your model-provider key and do not share either.
The original local dashboard used port 18789. To run a foreground gateway on that port:
openclaw gateway --port 18789Use the dashboard link or authentication method printed by your installed version. Keep the control interface local unless you deliberately configure authenticated remote access.
Configure browser access
The older interface used Config, Browser. The original example specified a CDP endpoint of http://127.0.0.1:18792, a Chrome profile, and the executable path C:\Program Files\Google\Chrome\Application\chrome.exe. An unused port alone does not start a browser: the endpoint must match a supported, running browser profile.
Configure handshake and connection timeouts for your environment. The original examples used 3–10 seconds. Check the current browser documentation because managed profiles and extension-based connections may use different settings.
Test a simple request in the dashboard or Telegram before a complex task. Limit access to accounts and files the task needs, and inspect what the agent proposes when a website requests a consequential action.
Adapted from the original Chinese article, published on February 1, 2026.
Comments NOTHING