wsbx
Docker-Sandboxes-style microVMs on Apple Silicon macOS. Each sandbox is a real Linux VM (Virtualization.framework via vfkit), booted from any OCI image with a Kata Containers kernel.
brew install vfkit e2fsprogs go # go: builds the static guest tunnel helper
cargo install --path . # or grab wsbx-*-darwin-arm64.tar.gz from a release (push a v* tag to build one)
wsbx claude # Claude Code in a VM, current dir mounted at the same path
wsbx codex | wsbx gemini # other agents (Docker's sandbox-templates images)
wsbx pi # pi coding agent, same version + config as on the host
wsbx run alpine # bare VM, interactive shell
wsbx run -w . ubuntu -- make test # one-off command, exit code propagated
wsbx run -p 8000 alpine # host's localhost:8000 reachable as localhost:8000 in the VM
wsbx exec NAME # shell in a running sandbox (alias: sh), as the image's user
wsbx exec -u root NAME -- apk add git # command, exit code propagated
wsbx forward NAME 3000 8080:80 # sandbox's localhost:3000 on the Mac's localhost:3000 (until Ctrl-C; alias fwd)
wsbx ps # sandboxes: status, CPU, memory, net/block I/O, PIDs, disk, image (alias: ls)
wsbx ps -f # same, refreshed every second like `docker stats`
wsbx rm NAME... # delete stopped sandboxes; --disk deletes only the disk (fresh on next start)
wsbx pull IMAGE
wsbx gc # delete cached images no sandbox uses
wsbx kernel [--update | --version TAG] # latest Kata release by default, sha256-verified
Defaults for --cpus/--memory come from the host (half the cores, a quarter of RAM up to 8 GiB). Image tags
are re-resolved after 24 h (cached copy used when offline); the superseded image is deleted automatically. WSBX_KATA_VERSION pins the kernel release.
Options: --name --workspace --cpus --memory --disk --env K[=V] --host-port/-p PORT --fresh --pull --kernel.
exec: host → vfkit unix socket (exec.sock) → vsock → guest helper. It runs in the workspace dir with the sandbox's env.
You get a pty (with resize) when stdin and stdout are terminals, plain pipes otherwise, so wsbx exec x -- cat f > out is byte-exact.
forward: host 127.0.0.1:HOSTPORT → exec.sock → vsock → guest helper → guest localhost:PORT, one exec-style session
per connection. Works for services bound only to the guest's loopback, on running sandboxes, no restart. 0:PORT picks a free host port.
Host tunnel (-p PORT): guest localhost:PORT → vsock → vfkit → unix socket → host localhost:PORT.
Nothing listens on a host network interface, and only the listed ports are reachable.
pi: installs @earendil-works/pi-coding-agent at the host's pi --version (reinstalls when it changes).
On every start it copies these from ~/.pi/agent with symlinks resolved: settings, models, keybindings, mcp, AGENTS/CLAUDE/SYSTEM md files,
extensions, skills, prompts, themes. They're linked into the guest's own ~/.pi/agent, so host edits
win and in-VM edits are discarded. npm/, bin/ (macOS binaries) and sessions stay in the VM. auth.json is only
copied with --pi-auth. Local baseUrls in models.json (127.0.0.1/localhost) are tunneled automatically.
Agent API keys (ANTHROPIC_API_KEY, OPENAI_API_KEY, …) are only forwarded with --keys. Code in the VM can read them.
Sandboxes persist per name (disk under ~/.wsbx/sandboxes); --fresh rebuilds. See PLAN.md for design.
Security model
The VM is the boundary. The guest (including a prompt-injected agent or a malicious image) is assumed hostile.
- Control share (
/.wsbx/host): the guest can write to it, so it's recreated empty on every start while the VM is stopped and deleted after exit. Guest outputs are read only as regular files with a size cap; persistent state lives in the host-onlymeta.json. - vfkit arguments: paths containing
,"or control characters are refused (vfkit splits options on commas). Workspaces overlapping~/.wsbxare refused. - Registry: references are validated before any request, digest pins are enforced, and every blob/manifest is sha256-verified. Auth realms must be https. For public registries, connections to loopback/private/link-local addresses are refused (token realms and redirects included).
- pi config: allowlisted items only. Symlinks are followed only into
~/.pi/agentor~/.agents/skills, with no cycles and no special files, capped at 256 MB. - exec: sessions run as root before dropping to the user, so the guest helper accepts them only from the host's vsock CID, not from guest processes via vsock loopback.
- forward: listens on the host's loopback only. Browsers don't isolate
localhostcookies by port, so a forwarded guest page can read non-HttpOnly cookies your other localhost apps set. Forward only what you need. - Host tunnel: only the ports you pass with
-p(or the localbaseUrls in pi'smodels.json), over vsock.
Remaining risks:
- The workspace is read-write. A guest can plant files that run later on the host (
.git/hooks,.envrc, IDE tasks, package scripts). Review changes before running them on the host. - Egress is unrestricted NAT. Anything the VM holds (keys passed with
--keys, API keys in pi'smodels.json,--pi-auth) can be exfiltrated. A host-side credential-injecting egress proxy would fix this; it isn't built yet. - A kernel panic before the root disk mounts (e.g. a corrupt
disk.img) still restart-loops; panics after boot stop the VM (exit 1). - Image layers are unpacked on the host into one tar, so a decompression bomb can fill the disk.