OpenClaw

OpenClaw is a program that you install yourself on a computer. It accepts instructions in plain language, for example: “Cancel on Anna and move the appointment to Friday.” It then acts largely on its own, rather than merely giving a reply. To do this, it can send messages, write to the calendar, open files, or visit websites. The source code is publicly viewable, so anyone may read it, modify it, and pass it on. That is exactly what distinguishes OpenClaw from assistants made by large corporations, which run as closed services on someone else’s servers.

Why an assistant on your own computer is causing a stir

Most well-known AI assistants are chat windows. You ask a question and get text back. OpenClaw goes a step further: it is allowed to actually do things. Experts call such systems agents, meaning programs that plan and carry out their own action steps. It is precisely this leap from “answering” to “acting” that is currently regarded as the most exciting area of the AI industry.

Then there is the aspect of control. Because the program runs locally, a lot of data stays on your own device. Tinkerers, developers, and privacy-conscious companies find that appealing. Within a few weeks, the project gathered tens of thousands of stars on the developer platform GitHub, i.e. public bookmarks from interested users. For a hobby project with no advertising budget, that is unusually a lot.

The downside became just as quickly apparent. A program that is allowed to send messages and read files is an attractive target for attackers. Security researchers found open access points and poorly secured settings in earlier versions. OpenClaw thus became a textbook example that action-capable AI creates new risks.

Router, tools, and memories: the architecture

OpenClaw does not think for itself. It sends your instruction to a large language model, i.e. to an AI that has learned from huge amounts of text how to continue language. Which model that is can be configured. OpenClaw is thus more like the manager than the mind: it organizes, delegates, and executes.

To make acting possible, the program has so-called tools. A tool is a fixed function, such as “send message” or “read file.” The language model decides which tool fits the task, and OpenClaw calls it. A memory store also keeps track of what has already been discussed. That way, the assistant still knows next time who Anna is.

OpenClaw is usually operated via ordinary messengers. You write to the assistant like a contact in WhatsApp or Telegram, while the program runs at home on a small server. This is also where the greatest danger lies: if a message from someone else contains instructions, the model may mistake it for a genuine task. Experts call this prompt injection, meaning the smuggling in of commands via seemingly harmless text.

OpenClaw in the headlines and in everyday life

In tech news, the name usually appears in two contexts. First, as proof that open-source projects can keep up with products from major providers. Second, as a cautionary example in reports about security vulnerabilities. The backstory is also often mentioned: the project was initially named differently and was renamed after trademark disputes.

In practical use, it is mainly used by technically savvy people. Typical tasks are sorting emails, coordinating appointments, maintaining shopping lists, or summarizing notes. Anyone who owns a smart home also uses it to control lights and heating. However, installation requires basic knowledge of servers and the command line.

For investors, OpenClaw is interesting as a signal, not as a product. It shows where the industry is heading: away from chat windows, toward software that completes tasks. A common misconception, by the way, is that OpenClaw is itself an AI. It is a shell that makes other models usable – and that is exactly where its appeal lies.