Over the past few months I’ve been experimenting with AI coding agents like Claude Code and have been blown away by what they can do with a well‑defined spec. At the same time, I’ve been hesitant to point them at my main codebase because I don’t fully trust them on my local machine. To keep things safe, I’ve been spinning up a separate VM whenever I need to run an agent-driven task, then tearing it down when I’m done. That workaround has let me customise agents and hooks while keeping my projects isolated — but it’s clunky and not exactly cost‑effective. This experience has led me to explore an idea I’m calling SentryForge: a secure, isolated sandbox where AI coding agents can run autonomously without exposing your source code or proprietary data. It’s still very early days — I’m trying to figure out what would make such a system trustworthy and useful. I’d love to hear from anyone who’s wrestled with similar concerns. What features would make you comfortable letting an AI agent run through your project? Do you see autonomous AI coding as part of your workflow in the near future? If you’re interested in shaping this concept, I’ve set up a waitlist (with some free runtime hours once there’s a beta): [https://waitlister.me/p/sentryforge](https://waitlister.me/p/sentryforge) Thanks for any feedback!