StuntDouble: eBPF sandbox & time-travel for local AI agents

👋 Hey Product Hunt! I'm Rohan, the creator of StuntDouble. A few weeks ago, I was running an autonomous AI coding agent on my laptop. It hallucinated, got confused, and almost dropped a local Postgres database I was actively using for another project. I realized that relying on "prompt engineering" (telling the AI "please don't delete my files") isn't real security. These agents have full root access to our local filesystems and networks. So I built StuntDouble. It’s a zero-trust wrapper that acts as a bodyguard for your local AI tools. Here’s the cool tech behind it: 🛡️ eBPF Kernel Hooks: StuntDouble injects raw C-code into your Linux kernel to monitor every outbound network packet the AI sends. If it tries to hit a destructive port (like Postgres or Mongo), the kernel silently drops the packet before it even leaves the virtual interface. The AI thinks it succeeded, but your host is untouched. ⏪ Zero-Copy Time Travel: Before the agent spins up, StuntDouble uses Git's low-level core object database (git write-tree) to take an instantaneous, cryptographic snapshot of your working directory. If the AI messes up your code, you just type stuntdouble rewind and your filesystem instantly reverts—without polluting your commit history! 🌐 StuntNet Swarm: You can spin up an entire fleet of agents (stuntdouble swarm coder tester) isolated inside a custom Docker bridge network that hard-blocks external internet access. It works out of the box with Aider, Claude Code, and Ollama. I built this entirely in Go, and it’s 100% open source. I'd love to hear your thoughts on using eBPF and kernel-level intercepts for AI safety. I’ll be hanging out in the comments all day to answer any technical questions! Cheers! 🍻