OpenClaw: the autonomous AI agent transforming your terminal

OpenClaw: A Revolution in AI Agent Autonomy

OpenClaw has become the most popular AI agent since its launch at the end of 2025. Created by Austrian engineer Peter Steinberger, who has since joined OpenAI, this agent embodies the promise of artificial intelligence capable of acting autonomously on your machine. Unlike attempts by Google and other companies, OpenClaw offers total autonomy without safeguards, raising security concerns. However, its open-source nature shifts the responsibility onto the user, who must know how to install and configure it correctly.

The OpenClaw AI agent stands out for its ability to automate routine tasks and make decisions autonomously. This simulated intelligence has been showcased by many tech giants, but none have succeeded in delivering a stable product. The main reason is that total autonomy without safeguards can lead to significant security risks that large companies are not willing to assume. The open-source nature of the project allows OpenClaw to exist, as the responsibility for its use lies with the person who installs it.

Although we do not recommend using OpenClaw as a professional agent, experimenting with it can provide insight into the future capabilities of unrestricted AI agents. However, it is essential to know how to install and configure it correctly to avoid any potential issues.

Choosing the Right Device for OpenClaw

The question of hosting OpenClaw is crucial. For personal and occasional use, a local installation may suffice. However, to isolate the AI and secure your data, a virtual machine running Unix is recommended. For more intensive use, a Mac mini or a mini PC can be considered, offering a good performance-to-price ratio. Finally, for a robust solution with no physical maintenance, a VPS server is ideal, although it requires networking and security skills.

For occasional experimentation in a personal setting, a local installation on your everyday machine is more than sufficient. However, if you want to isolate the AI and prevent it from accessing your system and files, you can set up a virtual machine with a Unix environment. The agent will be more comfortable on a Unix system.

On the other hand, if you genuinely want to industrialize the use of OpenClaw for personal or professional purposes, dedicating a machine becomes essential. On one hand, with a dedicated environment, the security risk is lower, as the agent does not have direct access to your files. On the other hand, if you want to use the agent through your preferred channels like WhatsApp or Telegram, the machine hosting the agent must run continuously.

In this case, you can opt for a local machine at home. Many experts recommend purchasing a Mac mini, both for its Unix environment (macOS), its physical compactness, its hardware designed for AI, and for the currently good performance-to-price ratio at Apple (699 euros). For a lower cost, a mini PC (300-400 euros) can also be considered.

The most robust solution is to use a virtualized server (VPS). This runs in the cloud without any physical maintenance on your part and allows you to truly control the execution environment of your agent. However, you will need networking and security knowledge to configure a "clean" installation on your VPS, which is a vital prerequisite before installing OpenClaw.

Installing OpenClaw on Windows

Installing OpenClaw in a Windows environment is simple and quick. Before you can install OpenClaw, you need to allow script execution from PowerShell. Start by opening PowerShell (press the Windows key and search for "Windows PowerShell," then click "Run as administrator"). Then enter the following command:

Set-ExecutionPolicy RemoteSigned -Scope CurrentUser

Once executed, the command will ask for your permission. Enter "Y" and then press Enter. Your PowerShell terminal is now authorized to execute scripts. Now let's proceed with the installation. In the same window, enter:

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

The installation of OpenClaw is now underway. The installation takes about 10 minutes, depending on your internet connection and hardware configuration. Once OpenClaw is installed, initiate the agent with the command:

openclaw onboard --install-daemon

OpenClaw will then ask you to accept the security risks associated with installing the agent on your machine. Accept them by selecting "Yes." You will then be asked for the type of configuration you want: default (QuickStart) or custom (Manual). Here, we choose QuickStart and will edit the settings later if needed. Next comes the question of selecting the model provider. This is a consideration that should not be overlooked.

Choosing the AI Model for OpenClaw

OpenClaw supports around thirty different providers by default: OpenAI, Anthropic, Google, xAI, MiniMax, Mistral… Laboratories from around the world are well represented. It is also possible to use a local model (via vLLM, for example). The model used by OpenClaw serves as the reasoning engine for your agent, so we recommend deciding on the provider based on the models available from them. Since an agent requires a true level of intelligence and reasoning, we advise using a high-performing model. With an SLM (less than 7 billion parameters), you may quickly find yourself limited in your more complex requests.

Your agent will primarily use system commands to execute your tasks, so we recommend choosing a provider that offers models with a very high score on the Terminal Bench benchmark (the model's ability to act in a terminal). Gemini 3.1, GPT-5.4, or Claude Opus 4.6 are excellent candidates. Using a local model should only be justified for reasons of privacy and cost. GLM-5 may suffice, but it will be much less effective than a proprietary model.

For this test, we will choose to use OpenAI as the provider with GPT-5.4 as the engine model, as the quality-to-price ratio is satisfactory.

Another point of attention: once the provider is selected, it is possible to use OAuth or the provider's API. OAuth uses your credentials associated with ChatGPT (via Codex) at OpenAI and those associated with Claude (via Claude Code) at Anthropic. While this method of authentication exists, it is often illegal under the service's terms of use. Anthropic bans many users each week who use this technique. Therefore, we recommend using an API key to eliminate any risk of banning. This highlights the importance of considering the quality-to-price ratio of the provider you will select.

Configuring External Access and Features of OpenClaw

The real uniqueness of OpenClaw lies in its external use, on your phone, via various channels: WhatsApp, Discord, Google Chat, Slack, iMessage, Teams… The easiest channel to configure is Telegram. To choose a channel, use the down and up arrows and press Enter.

Once Telegram is selected, go to your phone, in the Telegram app, to configure the connection with OpenClaw:

• Search for the bot "@BotFather."
• Click on "Create a New Bot."
• Give it a name.
• Then copy your bot's token.

Back in the OpenClaw terminal, paste your Telegram token. We then move on to configuring a search engine.

Search Engine, Skills, and Hooks

After configuring the external channel, OpenClaw asks for the default search engine to use. You can choose between Brave, Gemini, Grok, Kimi, and Perplexity. The simplest option is to use Gemini. So select Gemini and then provide your personal Gemini API key (Google offers a fairly generous free tier, sufficient for searching).

Once the search engine is configured, OpenClaw will ask you to configure skills, competencies it can install to perform complex tasks. More than twenty skills are offered by default. We recommend reading the description of each one carefully to know which will be a priority for you. In any case, OpenClaw will be able to install them for you later. In our case, we will install:

• video-frames to extract images from a video and use ffmpeg,
• summarize to summarize any type of files,
• openai-whisper to transcribe audio locally,
• nano-pdf to edit and read PDFs,
• mcporter to automatically add MCP servers from the web,
• clawhub to automatically install skills,
• and finally blogwatcher to monitor RSS feeds.

Next comes the configuration of hooks, automated scripts that trigger at certain key moments in the agent's life. Four hooks are offered by default.

• Boot-md loads a Markdown file at the agent's startup, allowing you to inject a custom system prompt at each launch.
• Bootstrap-extra-files works on the same principle by injecting additional files at boot, useful if your agent needs to access certain resources consistently.
• Command-logger records all commands executed by the agent.
• Session-memory saves the session context when you issue a /new or /reset command, allowing the agent to maintain persistent memory between sessions.

We recommend enabling all of these hooks, as their resource overhead is negligible compared to the flexibility they provide.

How to Use OpenClaw?

Once the installation is complete, your OpenClaw agent is operational. You can query it directly from your PC using the command

openclaw gateway

which opens a local web interface, or from your phone via the Telegram bot you just configured.