MicroVM for Agents
We are always happy about letting AI agents take over our computer and do all kinds of shit for us. But just consider it from another point of view: this is essentially allowing a remote entity to execute arbitrary commands on your system, which doesn’t sound as peaceful as it seems. For this reason, I have always been looking for any possible way to isolate the agent from my system. There are different solutions available on the market, varying from scripting with bwrap to various container- or disk-image-based solutions.
This means that I can’t re-use most of my host development environment; a re-installation is needed most of the time:
This means that I have to use a complicated bwrap command to confine the VM:
Agent-VM
The result of vibe coding is Agent-VM. I tried to combine my ideas when constructing a bwrap sandbox and the previously mentioned microVM solutions, and the result is that this tool, built with libkrun, supports:
- Construct a sandbox
/using the host/usrwith a read-only bind mount; - The minimal
/etcis constructed on the fly; - The current working directory is bind mounted to the sandbox;
- Allows various options to control the sandbox, like enabling/disabling network, passing certain host sockets to the sandbox, etc.
The configuration
The configuration should be quite self-explanatory. Here is a sample configuration file for the microVM:
~/.config/agent-vm/config.toml. One can have multiple different configuration files and specify the configuration file to use when starting the microVM using --profile <profile_name>. config.toml is the default one being looked for when no --profile is specified.version = 1
[vm]
cpus = 2
memory_mib = 2048
# Per guest-native tmpfs capacity limit, including [[tmpfs]] below.
# Actual use consumes VM RAM; this is not an aggregate limit or reservation.
tmp_mib = 256
[filesystem]
cwd = "rw"
home = "ephemeral"
# Source masks apply at every shared alias. Missing paths beneath a shared
# directory are rejected, because creating placeholders would modify the host.
mask_sources = ["~/.ssh", "~/.gnupg"]
# mask_targets = ["/data/private"]
# Relative source paths are resolved against this file's directory.
# [[mounts]]
# source = "~/datasets"
# target = "/data"
# mode = "ro"
# Empty guest-native storage; writes disappear when the VM exits.
# CLI --tmpfs entries append to this list. Each uses vm.tmp_mib above.
# UID/GID default to the invoking user's numeric IDs; mode defaults to 0700.
# Explicit guest IDs do not change host ownership or the workload identity.
# Root ownership with mode 0700 denies access to a non-root workload.
[[tmpfs]]
target = "/cache"
# uid = 1000
# gid = 1000
# mode = 0o750
# Optional temporary GnuPG home with an existing public keyring exposed read-only.
# Enable both entries together; do not also mask this source or guest subtree.
# UID/GID default to the invoking user's IDs; the explicit values are optional.
# [[tmpfs]]
# target = "~/.gnupg"
# uid = 1000
# gid = 1000
#
# [[mounts]]
# source = "~/.gnupg/pubring.kbx"
# target = "~/.gnupg/pubring.kbx"
# mode = "ro"
# Forward an existing filesystem Unix stream socket. Relative sources use this
# file's directory; ~ expands to the invoking user's home, as for mounts.
# CLI --socket entries append to these. Targets must be absolute and writable.
# Missing parent directories are created as guest root, then assigned to the
# guest user before the relay drops privileges and creates a mode-0600 socket.
# Existing parent ownership is unchanged; existing destinations are rejected.
# [[sockets]]
# source = "~/service.sock"
# target = "/run/service/client.sock"
[environment]
# TERM, LANG and LC_ALL are inherited by default when present.
# Add only variables that the workload needs. Missing explicit names are errors.
inherit = []
[environment.set]
EDITOR = "vi"
[network]
mode = "none" # or "passt" to enable host network passthrough
# Publishing requires mode = "passt". Default bind address is loopback.
publish = []
# publish = ["127.0.0.1:8080:8080/tcp"]
[ssh_agent]
# Alias: host SSH_AUTH_SOCK -> /run/user/<uid>/ssh-agent.socket, with guest env.
# --no-ssh-agent disables this alias only, not explicit [[sockets]] entries.
# When disabled, environment.set may specify SSH_AUTH_SOCK for a custom forward.
enabled = false
# Optional filtered D-Bus forwarding; each bus is independently disabled by default.
# Arguments are literal xdg-dbus-proxy per-bus options; --filter is always added.
# No shell expansion. Empty args exposes only the proxy's baseline bus operations.
[dbus.user]
enabled = false
args = ["--talk=org.freedesktop.Notifications"]
# address = "unix:path=/run/user/1000/bus"
# Otherwise uses host DBUS_SESSION_BUS_ADDRESS, then $XDG_RUNTIME_DIR/bus.
[dbus.system]
enabled = false
args = ["--talk=org.freedesktop.UPower"]
# address = "unix:path=/run/dbus/system_bus_socket"
# Otherwise uses host DBUS_SYSTEM_BUS_ADDRESS, then the standard system socket.And extra configuration can be applied via cmdline like agent-vm --mask .git -- some_coding_agent to hide the .git directory from being uploaded by accident.
Installation
Currently packaging is only available for Fedora on Copr, but it should be easy to build from source on other distributions. The build process is quite simple. Just run:
git clone https://github.com/karuboniru/agent-vm.git
cd agent-vm
cmake -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo
cmake --install build --prefix ~/.local