omabox and ai-jail
ai-jail fences the agent in. omabox keeps its windows out.
omabox keeps an agent's windows off your desktop; ai-jail fences what the agent can read, write and reach. They stack. omabox on its own is not a security boundary: a box shares your kernel, your GPU and, unless you isolate it, your network.
- ~/code/app, read-write
- its model's API
- ~/.ssh, ~/.aws, ~/.gnupg
- your shell's tokens
- your LAN
0 windows · 0 focus taken · 0 prompts
ai-jail decides what the agent can touch. omabox decides where it draws.
Two questions, two tools.
ai-jail, by Fabio Akita (AkitaOnRails),
is the sandbox we point people to. It runs an agent inside bubblewrap, Landlock and seccomp on Linux, and
sandbox-exec on macOS. It does carefully the one thing omabox doesn't do: limit what the agent can read,
write and reach.
ai-jail answers: what can it touch?
A fence around the agent
- Your project read-write at its real path,
/usrread-only - A fresh home: no
~/.ssh,~/.aws,~/.gnupgor browser profiles - The environment cut to an allowlist, so your tokens stay in your shell
- The network cut to what you allow: since ai-jail 2.8, a bare
ai-jail claudereaches its model's API and nothing else
omabox answers: where does it draw?
A desktop of its own
- Its own screen, pointer and keyboard
- Its own session bus, tray, notifications and keyring
- Your cursor, focus and workspaces untouched
- Not a security boundary: that part is ai-jail's
Two recipes, both tested.
Neither needs a change to ai-jail.
A jailed app, drawn in a box
A contributor's branch, an app you just downloaded: run it in ai-jail, with the box's screen as its only display.
ai-jail keeps it out of your home, keys and network; its window opens in the box, where omabox shot and
peek see it.
$ omabox up $ eval "$(omabox env)" # Wayland clients now draw in the box $ ai-jail --gpu --rw-map "$WAYLAND_DISPLAY" \ --env WAYLAND_DISPLAY -- ./build/app
Tested with ai-jail 2.10.0 and 2.12.0 and omabox 1.0 on Linux, and before that with ai-jail 2.2.1, 2.6.2 and 2.8.1.
A jailed agent, driving boxes of its own
A jail can't start a box itself (its seccomp filter refuses new namespaces), so its omabox hands each
command to a broker outside. Turn it on once, add the lines it prints, then start your agent in ai-jail as usual.
$ omabox broker on # prints the lines to add to ~/.ai-jail $ ai-jail claude
Tested with ai-jail 2.10.0 and 2.12.0 and omabox 1.0 on Linux, and before that with ai-jail 2.2.1, 2.6.2 and 2.8.1.
omabox broker off turns it off.
What a jailed agent can and can't do.
Driving a box is running code in it, so the broker is what keeps a box from being a way out of the jail. It reads the jail's limits from the jail's own command line, which nothing inside can change, and gives each box no more than the jail has.
In the jail, as outside
up,run,shot,keys,click,waitand the restrecord: the video, its marks and captions land in the jail's project, written by the agent's ownomaboxports, for its own boxes: a port held by anything else is said to be taken, without naming who holds it- git in a box started from a worktree
Not from a jail
omabox host, which runs on your real desktoppeekand interactive boxes, which open on your screen- The guard, saves,
clipandconfigchanges; which host ports you protect isn't shown to it - Boxes you or another jail started
None when the jail has none: its boxes start --net isolated. With ai-jail 2.8's
default, the agent reaches its model's API and its boxes still get no network.
Only the jail's own project goes into a box, and what the jail hides in it (its
.ai-jail included) stays hidden there too.
Only the boxes that jail started, and they go down when the jail exits.
Upgrading from 0.x to 1.0? Restart agents running in ai-jail: a jail started before the upgrade keeps its old relay. And omabox is still not a security boundary. A box keeps an agent's windows off your desktop, not code written to escape: for code you don't trust, use ai-jail, or a VM. The rule for jailed agents and the test behind it are S9 in SECURITY.md.
Give every agent a desktop of its own.
git clone https://github.com/diogochaves/omabox && cd omabox && ./install.shToday on Omarchy 4 · Hyprland 0.56+ · AMD, Intel or NVIDIA · MIT