The Lobster That Took Over the Internet
OpenClaw, the open-source AI agent with the lobster mascot, has amassed over 150,000 GitHub stars and 300,000 users since launching in late 2025. It connects to any large language model, operates through virtually any messaging platform, and actually does things: reads email, manages calendars, runs code, and remembers context across conversations.
It started as a hobby project by Peter Steinberger before OpenAI acquired it through an acqui-hire. Peter is now working with OpenAI, and plans to bring agents to the mainstream.
What OpenClaw represents matters more than what it does. It’s an early preview of the agentic harness, a framework where any AI brain plugs in and operates like a digital employee.
A note on security, from FPOV’s Corey White, CCII:
OpenClaw is definitely an interesting tool, one that should be examined and tested. However, it is not quite ready for primetime. The vulnerabilities in OpenClaw have been well documented and handing over administrative level privilege to an automated system that can act on your behalf is not recommended, at least not yet. OpenClaw, and systems like it, are granted wide privileges so a compromised agent could have access to sensitive information that could be exposed or stolen by a bad actor or inadvertently misused by the system.
In one extreme case, a Meta AI researcher allowed OpenClaw access to her email and the system proceeded to delete her email inbox. In another experiment, a researcher was able to trick the system into sending a summary email to an “attacker” thereby leaking the last five emails the researcher had sent. There are other cases like this.
It is important not to dismiss OpenClaw, in spite of its risks. You must instead take precautions. You can utilize a tool such as IronClaw which acts like OpenClaw but runs in a sandbox environment without having privileged access to an entire system. Also, you can create your own sandbox for testing as our very own Jake Holmquest has done…..
From Future Point of View’s R&D team :
“For our own OpenClaw instance, we started by sandboxing the system on a dedicated laptop running a Linux environment. That isolation means it runs 24/7 but only has access to exactly what we grant it.
Our first use case was a digital research intern. We gave it its own Gmail account, connected a ChatGPT Pro subscription to power it, and communicate with it primarily through email. Though we do have telegram configured just so I can ensure its operating well throughout the day and know what its working on.
What we’ve found is that while ChatGPT, Claude, or Copilot can do similar research, there’s real value in staying inside your inbox rather than switching platforms. You forward something, it takes action, the output comes back. The workflow stays intact.
Right now we’re using it primarily for news monitoring and research. We’re also beginning to test it in a light project management capacity, though we’re moving deliberately. Expanding an agent’s role inside your organization is the kind of thing you want to grow into carefully, not rush.”