Ask HN: What are agent sandboxes missing?
I'm building an agent sandbox platform with opinionated customizable templates for general purpose agents using the OpenCode SDK.
The idea is to provide a great CLI experience so it's easy to use an AI client like Claude Code or Codex to launch sandboxes, fetch/retrieve data and prompts, review generated outputs, and manage queues for ask-user questions.
I also plan to add an AI gateway layer, similar to Tailscale Aperture(not affiliated but one of their engineers gave a great talk at AI Engineer it's on YT) for monitoring network access, distributing secrets, and controlling what services/systems a sandbox has access to.
For people building agent workflows, I'd love to hear feedback if you've tried other sandbox solutions.
What was missing and what was the hard part? What I disliked about the Claude sandbox is that their IPs are blocked by many platforms, which makes webfetch etc. nearly impossible. I had to write a skill to avoid that. Best example is reddit that blocks all access for LLMs. I mean, it's surely understandable why those paltforms do it. But for your own purpose like research and can be a bit annoying. Docker sandboxes do that. They take quite bit of resources though, almost same as a standard VM. I need to mount directories or remove after creation easily, like I do with containers. Gateway would be good too. Firewall needs more functionality. You can use micro VMs like Firecraker if you're rolling out your own infra. If you're on AWS, they also have a new managed Lambda MicroVM service. Can you share more on what functionality you'd like from firewall? Feel free to drop me a message on oryx1729@protonmail.com if you'd like an early preview of my product.