Safer Agentic AI? 🧠 Full Docker + Hermes Agent Tutorial (Desktop)

Callum (Wanderer) explains the security risks of running AI agents with unrestricted access to your computer, including a real example where Meta's privacy head accidentally deleted emails using Claude. The core problem is that AI agents can read and execute code directly on your machine, potentially exposing sensitive information like API keys, passwords, personal notes, and company strategies.

The solution is Docker, which creates an isolated sandbox environment where agents can execute code without affecting the host system. The video demonstrates this by first showing Hermes reading all files in a test folder containing sensitive information (secrets, server details, personal therapy notes, business strategy), then reconfiguring Hermes to run inside Docker where it cannot access those same files.

The setup process involves installing Docker Desktop, modifying Hermes' configuration file to set the working directory to /workspace, and adding Docker volume mounts that link specific folders on your computer to the container. For example, mounting /Users/username/hermes-workspace to /workspace inside the container allows selective file access. The speaker demonstrates this by having Hermes create files that appear both in the Docker container and the linked desktop folder.

The video emphasizes the distinction between access and execution: agents should have access to only the files you explicitly choose (through mounted volumes), but execution of code should happen only inside the container. This prevents malicious code, accidental deletions, and data theft. However, the speaker notes this isn't 100% foolproof—custom skills may still execute on the host computer—so users must still be selective about which tools and skills they install.

An advanced option is running the entire Hermes instance inside Docker and accessing it remotely via a gateway, providing complete isolation but with some usability tradeoffs. The speaker recommends the basic Docker execution approach for everyday use while reserving full containerization for higher-risk scenarios.