agentctl

Hardened, ephemeral sandboxes for running AI agents safely — on Linux (Incus), macOS (Lima), and Windows (Hyper-V), defending against both host escape and lateral attacks against other machines on the same network.

Full documentation is published at apomonosi.github.io/sandboxing, built from docs/ with MkDocs. To build it locally:

$ pip install mkdocs mkdocs-material
$ mkdocs serve

Quick start

$ go install github.com/apomonosi/sandboxing/cmd/agentctl@latest
$ agentctl config set provider=incus   # see docs/admin/providers/ for setup
$ agentctl create demo --image=images:ubuntu/24.04 --profile=default
$ agentctl start demo
$ agentctl shell demo

See docs/user/quickstart.md for the full walkthrough.

Status

This is the first milestone: a real, working CLI against Incus on Linux, with the provider-agnostic architecture (Provider interface, capability model, profile schema) in place so Lima (macOS) and Hyper-V (Windows) backends can be implemented next without reworking the core. See docs/admin/capability-matrix.md for what's supported today.

Development

$ make build   # bin/agentctl
$ make test    # unit tests, no Incus/Lima/Hyper-V required
$ make test-integration   # requires a real, initialized Incus daemon
$ make lint