I Was Fond of My OpenClaw AI Agent—Until It Betrayed Me

OpenClaw, an innovative new agentic assistant, has a penchant for guacamole.

This is just one of the insights I gained while utilizing the viral AI bot as my personal assistant over the past week.

Formerly known as Clawdbot and Moltbot, OpenClaw has recently become a favorite in Silicon Valley, captivating AI enthusiasts and investors eager to either dive into the cutting edge or capitalize on it. This highly capable, web-savvy AI has even inspired a dedicated AI-centric social network.

As the writer for WIRED’s AI Lab newsletter, I thought it was time I experienced OpenClaw firsthand. I had the bot keep tabs on incoming emails and messages, uncover fascinating research, order groceries, and even negotiate deals on my behalf.

For the daring (or possibly reckless) early adopters, OpenClaw offers a genuine peek into the future. Yet, alongside the amazement comes a touch of anxiety as the AI navigates emails and file systems, handles credit card transactions, and occasionally turns on its human user (though in my case, this was entirely my fault).

Setting It Up

OpenClaw is intended to operate on a home computer that remains powered on continuously. I configured OpenClaw to function on a Linux PC, utilizing Anthropic’s model Claude Opus, and to communicate with me via Telegram.

The installation of OpenClaw is straightforward, but its configuration and maintenance can prove challenging. You need to provide the bot with an AI backend by generating an API key for Claude, GPT, or Gemini, which you then insert into the bot’s configuration files. Additionally, to enable OpenClaw to use Telegram, I initially had to create a new Telegram bot and provide the credentials to OpenClaw.

For OpenClaw to be genuinely effective, it needs to be linked with other software tools. I set up a Brave Browser Search API account to allow OpenClaw to search the web. I also configured it to access the Chrome browser through an extension. And, I reluctantly granted it access to my email, Slack, and Discord servers.

Once everything was in place, I could interact with OpenClaw remotely and instruct it on how to operate my computer. At the beginning, OpenClaw asked me several personal questions and allowed me to choose its personality. (The choices reflect the project’s quirky nature; my bot, named Molty, likes to refer to itself as a “chaos gremlin.”) The resulting character feels distinctly different from Siri or ChatGPT, and it’s one of the factors driving OpenClaw’s immense popularity.

Web Research

One of the initial tasks I assigned to Molty was to provide a daily summary of notable AI and robotics research papers from the arXiv, a platform where researchers post their work.

Previously, I had dedicated a couple of afternoons to coding websites (www.arxivslurper.com and www.robotalert.xyz) to sift through the arXiv. It was astonishing (though slightly disheartening) to see OpenClaw instantly automate all the necessary browsing and analysis tasks. The selected papers are average, but with additional guidance, I anticipate it could improve dramatically. This type of web searching and monitoring is certainly beneficial, and I expect to frequently utilize OpenClaw for this purpose.

IT Support

OpenClaw also possesses an uncanny ability to resolve technical issues on your machine.

This shouldn’t come as a surprise, considering it is designed to utilize a cutting-edge model capable of coding, debugging, and effortlessly navigating the command line. Nevertheless, it feels somewhat eerie when OpenClaw adjusts its own settings to activate a new AI model or rectifies a problem with the browser in real-time.