The video demonstrates building a personal AI assistant named Clank using the Hermes AI agent harness on a cloud-hosted VPS, showcasing its customizable features, memory system, and integrations while emphasizing security best practices and cost-effective model choices. It concludes with a hacking experiment revealing that, despite some vulnerabilities, layered defenses make such assistants reasonably secure for practical use, encouraging viewers to explore creating their own.
The video begins by referencing the movie “Her,” where the protagonist installs a personal AI assistant, sparking curiosity about the realism and safety of having such assistants today. The creator introduces the concept of building a personal AI assistant using cloud-hosted virtual private servers (VPS) rather than expensive home labs. The main technologies involved are Hermes, an AI agent harness, and large language models (LLMs) that serve as the brain. The assistant can be interacted with through various integrations like browsers, Discord, Telegram, or email, and Hermes offers features like a “soul” document defining personality, memory stored in markdown files, skills for specific tasks, and a curator that auto-improves the assistant over time.
The setup process is demonstrated using Hostinger’s VPS service, sponsored in the video, where Hermes is installed either via a managed service or manually for more control. The presenter connects Hermes to an inference provider, specifically News Portal, to power the AI brain with large language models. Once connected, the assistant—named Clank—is tested through Discord, showcasing its ability to perform tasks like checking the weather and learning from interactions. The assistant continuously saves memories and can auto-generate or improve skills, making it a dynamic and personalized tool that runs persistently in the cloud, unlike local solutions that stop when the device is off.
The video delves into the internal workings of Hermes, highlighting features like scheduled tasks (cron jobs) for automating workflows, skills that bundle instructions for specific tasks (e.g., recipe planning), and a simple but effective memory system using markdown files. The presenter emphasizes security best practices such as applying the principle of least privilege when granting API keys, ensuring the assistant only has minimal necessary access. The modular and transparent nature of Hermes allows users to inspect, modify, or even have the assistant modify its own code and skills, providing a highly customizable experience.
Cost considerations are discussed, with the presenter sharing their experience of initially using expensive models like Sonnet, which quickly became costly, then switching to more affordable yet capable models like DeepSeek V4 Flash. These cheaper models still offer strong performance for personal assistant tasks and include safeguards against prompt injection attacks. The video explains the layered security approach: using modern, fine-tuned models resistant to prompt injection, input-output filtering by the harness, minimal privilege credentials, and isolating the assistant on a secure VPS rather than a personal machine containing sensitive data.
Finally, the presenter conducts a hacking experiment by exposing the assistant to various prompt injection attacks via a public email inbox. Despite 50 attempts, only one minor data leak occurred, demonstrating that while prompt injection is still possible, it is much harder than expected. The assistant effectively blocked attempts to leak sensitive API keys and other critical information. The takeaway is that personal AI assistants like Hermes are becoming secure enough for practical use if users apply defense-in-depth strategies and carefully manage sensitive data and access. The video encourages viewers to experiment with building their own assistants, highlighting the growing feasibility and power of these tools.