An open-source project called Clawdbot has become one of the most talked-about pieces of AI software of January 2026. Rather than a chatbot that answers questions, it is a personal agent that runs on the user’s own hardware, remembers context across weeks, and carries out tasks on the user’s behalf. Supporters have described it as “Claude with hands,” a reference to the Anthropic models it is most often paired with.
Clawdbot was created by Peter Steinberger, an Austrian developer previously known for the PSPDFKit software library. The core idea is to wrap a powerful language model in an always-on agent that has persistent memory, access to the user’s systems and the ability to pursue goals rather than respond to single prompts.
Self-hosted and persistent
The project’s distinguishing feature is that it is designed to be self-hosted. It can run on a Mac, a Linux or Windows machine, an old laptop, a Raspberry Pi or a rented virtual server. The agent itself does not depend on a central cloud service, although it calls out to model providers such as Anthropic or OpenAI for the language model behind it.
Its memory persists between sessions, so it retains earlier conversations, preferences and ongoing tasks instead of starting fresh each time. The agent can also initiate contact: it can send daily briefings, reminders or alerts through Slack, Telegram, WhatsApp and other messaging apps. With permission, it can read email, manage a calendar, automate web tasks through a browser, run code and scripts, and work directly with developer tools.
Rapid growth and real risks
The project’s GitHub repository accumulated tens of thousands of stars within weeks, and discussion spread across Reddit, Hacker News and X, where some users compared it to a real-world version of a science-fiction assistant. Others described buying a dedicated Mac mini or server just to run it in isolation, which says something about both its appeal and the caution it demands.
That caution is warranted. Because the agent can be granted deep access to a machine, security and privacy concerns are real. The standard advice from security-minded users is to run it in a separate or isolated environment rather than on a primary computer that holds personal data, since prompt injection (malicious instructions hidden in content the agent reads) or a misunderstood command could cause serious damage. Running the agent can also be expensive in API usage, particularly when large models are used for complex tasks.
A forced rename
Within days of its surge in popularity, the project announced it would be renamed Moltbot. The change followed a trademark request from Anthropic over the similarity between “Clawd” and “Claude,” and the developer chose a quick rebrand rather than a legal dispute. The project’s features, open-source license and autonomous capabilities were unchanged. The name was changed again shortly afterward to OpenClaw, which is how the project is known today.
Whatever it is called, the project is a clear marker of where consumer AI is heading: away from assistants that only talk and toward agents that act, with all the convenience and responsibility that implies.
