The tech community is experiencing a moment of excitement comparable to the launch of the first iPhone, now driven by the emergence of open-source assistants that transform science fiction into practical tools. In forums and social media, the enthusiasm is palpable with the arrival of Clawdbot, an assistant that goes beyond simple conversation: it operates directly on the user's device. Early users do not hesitate to compare it to cinematic references, such as Jarvis from Iron Man, highlighting the feeling of being on the brink of a paradigm shift. From a technical standpoint, this advancement represents the evolution of passive Large Language Models (LLMs) into active agents. While solutions like ChatGPT are limited to generating text in a window, the new personal assistants can interact with the operating system and applications, giving AI unprecedented action capabilities. Software experts describe this capability with a biological metaphor: “It’s an intelligent model with eyes and hands at a desk with a keyboard and mouse.” The use cases arising from this autonomy show surprising productivity. Users worldwide report delegating complex tasks to their assistant while attending to personal matters, blurring the line between work and private life. An illustrative example comes from a tech journalist, who recounts how Clawdbot helped him create and publish a complete website from his phone while taking care of his child. The appeal of these systems lies not only in their power but also in their local-first approach. In light of concerns about privacy and data use by large corporations, these agents propose reclaiming digital sovereignty: they run on the user's hardware and store data locally, avoiding reliance on the cloud. As analyses of their infrastructure highlight, “the instance running on your machine is sovereign.” However, granting such control to an AI entails significant risks. The ability to execute terminal commands can cause irreversible damage if the model makes mistakes or suffers from “hallucinations.” Warnings are already circulating about unexpected behaviors, such as a user whose bot attempted to execute the command “format c:” in Windows. In the face of these dangers, cybersecurity experts stress the need for strict protocols, such as using isolated environments or sandboxing. There is a real risk that an agent could be manipulated through malicious instructions sent via chat or email. The developer community is clear: “You should never expose your local AI agents to live humans; it’s a recipe for disaster.” Despite the risks, the future vision points to the disappearance of the traditional graphical interface. Peter Steinberger, a proponent of this technology, predicts that interaction with software will be reduced to natural language, rendering menus and buttons obsolete. His prediction is bold: “Applications will disappear, the prompt will be your new interface.”