I loved my OpenClaw AI agent, until it turned on me

OpenClaw, a powerful new proxy assistant, has a thing for guacamole.

This is one of the many things I discovered while using the virus artificial intelligence Bot as my personal assistant last week.

Formerly known as both CloudBot and MultiBotOpenClaw has recently become a darling of Silicon Valley, enchanting AI enthusiasts and investors eager to either embrace the bleeding edge or profit from it. The highly capable, web-savvy AI bot was an inspiration Its only (or mostly) AI social network..

As a writer for WIRED Artificial Intelligence Laboratory Newsletter, I thought I should take a chance and try OpenClaw myself. I asked the bot to monitor incoming emails and other messages, look for interesting research, order groceries, and even negotiate deals on my behalf.

For brave (or perhaps reckless) early adopters, OpenClaw seems like a legitimate glimpse into the future. But any sense of astonishment is accompanied by a tinge of horror as the AI ​​agent navigates through emails and filing systems, uses a credit card, and occasionally turns on its human user (although in my case, that change was entirely my fault).

How do I set it up

OpenClaw is designed to live on a home computer that is turned on all the time. I configured OpenClaw to run on a Linux computer, and for Anthropic’s model, Claude Opus, to reach out and talk to me via Telegram.

Installing OpenClaw is simple, but configuring it and keeping it running can be a headache. You need to give the bot an AI backend by creating an API key for Claude, GPT, or Gemini, which you paste into the bot’s configuration files. In order to get OpenClaw to use Telegram, I also had to do that first creates New Telegram bot, then give the bot credentials to OpenClaw.

For OpenClaw to be really useful, you need to connect it to other software tools. I created a Brave Browser Search API account to allow OpenClaw to search the web. I also configured it so that it can access the Chrome browser through an extension. And God help me, I gave him access to my email servers, Slack, and Discord.

Once all this is done, I can talk to OpenClaw from anywhere and tell it how to use my computer. First, OpenClaw asked me some personal questions and let me choose his character. (The choices reflect the project’s chaotic vibe; my bot, named Molty, likes to call himself a “chaos gremlin.”) The resulting persona looks very different from Siri or ChatGPT, which is one of the secrets of OpenClaw’s runaway popularity.

Web research

One of the first things I asked Molty to do was send out a daily roundup of interesting research papers on AI and robotics from arXiv, a platform where researchers upload their work.

I previously spent many afternoons programming websites (www.arxivslurper.com and www.robotalert.xyz) to search arXiv. It was amazing (though a little frustrating) to see OpenClaw instantly automate all the browsing and analysis work required. The papers she chooses are okay, but with more instruction I imagine they could improve a lot. This kind of web searching and monitoring is certainly useful, and I imagine I’ll be using OpenClaw a lot for this.

IT support

OpenClaw also has a miraculous, almost scary ability to fix technical problems on your device.

This should not be surprising, since it is designed to use a parametric model that is able to write and debug code and use the command line with ease. However, it is strange for OpenClaw to reconfigure its own settings to quickly load a new AI model or correct a browser issue.