cd /news/ai-agents/an-outer-sandbox-for-coding-agents · home › topics › ai-agents › article
[ARTICLE · art-172532] src=mo.town ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

An Outer Sandbox for Coding Agents

Developer 0xfe released lilbox, an open-source wrapper that runs coding agents inside an "outer sandbox" using Seatbelt on macOS and bubblewrap on Linux, so an agent that is granted a bad command can only reach what the profile allows. The default lilbox profile grants access to the project directory, login state, selected tools, caches and the network while keeping common credential files out of reach and filtering the inherited environment. The tool targets developers who find Docker images and VMs too heavy for quick projects but want to limit what an agent can read from their home directory or shell environment.

key takeaway

Run coding agents inside an outer sandbox — lilbox uses Seatbelt on macOS and bubblewrap on Linux to restrict the agent to the project directory, login state, selected tools and caches, while blocking common credential files and filtering the inherited environment.

read2 min views1 publishedNov 19, 2026

This is a post about mo.town/lilbox. I use a lot of coding tools, and I like trying out new models and using them for real work. But with many of them, it's very hard for me to be sure that they're not exposing the sensitive data in my home directory or shell environment.

What worries me most is a tool reading something it has no business reading, and then finding out later that the contents are with the model provider, or worse, shared with other third parties.

Coding tools have gotten pretty good at sandboxing the shell commands they run. The catch is that the program launching the shell can have access to a lot more than the shell itself.

Approval prompts help, but I know how carefully I read the first one and how carefully I read the twentieth.

I'd really like to decide up front what an agent can or cannot see.

Docker and VMs work great for this. For a quick project, though, putting together an image with my compiler, my tools, and a working agent login can be more work than the thing I sat down to do.

An outer sandbox #

So the (very simple) idea of an outer sandbox is to start the coding tool itself inside a dedicated lightweight sandbox. I can still use the agent's own approval system, but also be a bit more confident that if I approve the wrong command, it can only reach what I allowed into the outer sandbox.

lilbox #

I wrote a little wrapper for this called lilbox (mo.town/lilbox). It uses Seatbelt on macOS and bubblewrap on Linux, both lightweight OS-native sandboxing subsystems.

The default profile gives the agent access to the project directory, its login state, selected tools and caches, and the network. It keeps common credential files out of reach and filters the environment the agent inherits. But, what I like about it most is that I can build or edit a profile really quickly with an agent.

There are many agent isolation tools, but this one's mine. :-) Try it out.

A bit more on the rationale is at [Why Outer Sandboxes?](https://mo.town/lilbox/#why-title)

The source and more example profiles are on [GitHub](https://github.com/0xfe/lilbox).
── more in #ai-agents 4 stories · sorted by recency
── more on @lilbox 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/an-outer-sandbox-for…] indexed:0 read:2min 2026-11-19 · —