The Isolation section predated the #16 arc: it described the ACL and the host firewall, and none of what that audit shipped — port-isolated NICs (the actual sibling boundary), dns.mode=none, the pinned resolver. It also never mentioned that the contract is tested at all. State the contract layer by layer as it exists on main, and add a 'Measured, not claimed' section: the drill and doctor, what each does, which one mutates the host, and the standing score — 47/47 from a bare host (run 13, full teardown → rebuild → drill).
8.9 KiB
claudebox
A CLI to run headless, trust-less Claude Code in throwaway VMs. One command mints a fresh, network-isolated Incus box with Claude Code installed. The box is the product — you log in and work; destroying it loses nothing you didn't push.
Strictly creds-free. A box ships with everything installed and no credentials — no Claude token, no git PAT, nothing. You authenticate interactively inside the box. The tool never stores or injects a secret. That means there's nothing shared or committed, so it's safe for multiple operators out of the box.
The tool knows nothing about your projects. You just git clone inside a
box. A repo can ship an optional .claudebox/
runbook that Claude Code reads and acts on — there is no install step and no
host-run setup. See docs/claudebox-design.md for the
design rationale.
Install
curl -fsSL https://raw.githubusercontent.com/heavy-duty/claudebox/main/install.sh | bash
Installs the tree to ~/.local/share/claudebox and links claudebox onto your
PATH. Re-run any time to upgrade. (No git clone needed.)
One-time host setup (Ubuntu 24.04 / Debian 13)
~/.local/share/claudebox/host/setup-host.sh # run twice if it adds you to incus-admin (re-login between)
Idempotent. Installs Incus and creates the isolation stack: the claudenet NAT
bridge (sibling-name resolution off, resolver pinned to public upstreams —
BOX_DNS overrides), the claude-isolate ACL (drops all RFC1918/CGNAT/
link-local egress), the claude-dev profile (port-isolated NICs — boxes can't
reach each other), and firewall rules blocking instance → host. All rules
re-apply at boot via claudebox-firewall.service — no post-reboot ritual. If
the host lacks dnsmasq-base (Debian cloud images skip Recommends):
sudo apt-get install -y dnsmasq-base.
Quick start
claudebox new --name work # mint a fresh, creds-free box (~10 min cold)
claudebox shell work # enter as the claude user
Inside the box, authenticate as needed:
claude # then run /login — copy the URL (press c), open it
# in YOUR browser, paste the code back. No host CLI needed.
gh auth login # or drop a PAT in — your git credentials, your call
git clone https://github.com/you/project && cd project
claude # if the repo has .claudebox/, Claude reads it and sets up
Log in once, reuse via snapshots
Because every fresh box is creds-free, re-authenticating each time would be toil. Snapshot an authenticated box and clone from it instead:
claudebox snapshot work authed # checkpoint after you've logged in
claudebox new --name feature --from work/authed # clone the authed state into a new box
--from copies the whole box (Claude login, git creds, clones and all) while
preserving isolation. You can also claudebox new --name x --from work to clone
a box's live state, or roll a box back with claudebox restore work authed.
Forgotten what you called a checkpoint? claudebox info work prints the box's
snapshot labels and the --from line to clone one.
Commands
claudebox new --name <box> [--from <src>[/<snap>]] [--vm|--container] [--remote r]
claudebox list # list your boxes
claudebox info <box> # one box: state, IP, snapshot labels
claudebox shell <box> # enter as the claude user
claudebox exec <box> -- <cmd...> # run a command in the box
claudebox snapshot <box> [label] # checkpoint (label defaults to manual-<epoch>)
claudebox restore <box> <snap> # roll back to a snapshot
claudebox rename <box> <new> # rename a box (stop it first)
claudebox down <box> # stop (state kept; `start` resumes)
claudebox start <box> # start a stopped box
claudebox rm <box> [--force] # delete the box + its snapshots (asks first)
claudebox incus <box> -- <args...> # escape hatch: any incus command, box resolved
claudebox status # deprecated alias for `list`
claudebox help [<command>] # full help, or one command's page
Every command takes --help, and options come after the command
(claudebox list --json). Exit status: 0 ok, 1 it went wrong, 2 you asked
wrong.
new fresh-launches from cloud-init, or with --from clones an existing box or
snapshot. VM mode (--vm, the default where /dev/kvm exists) is the trust-less
target; container mode (auto-fallback, security.nesting=true) is for hosts
without nested virt — weaker isolation, dev/test only.
Boxes are just Incus instances
A box is an ordinary Incus instance tagged user.claudebox=1. claudebox wraps
the box lifecycle and the isolation model — not all of Incus. It owns a command
when it must enforce something Incus can't see: that tag (it will not stop,
rename or delete an instance it didn't mint), the isolation stack, or the
creds-free snapshot workflow. For everything else, there's the door:
claudebox incus work -- config show # instance name appended
claudebox incus work -- file push x.tar {}/tmp/ # or placed with {}
The box is resolved and tag-checked; the rest is passed to incus verbatim, and
the command is echoed before it runs. If it can move the box off the isolation
stack (profile, network, device, security.*), claudebox warns and proceeds —
the trust boundary is then yours to keep. See
docs/claudebox-design.md for the rule and why the
command surface is a table.
Isolation
The contract: a box reaches the public internet and nothing else. Not the host, not your LAN, not another box, not even another box's name. What enforces it, layer by layer:
- Dedicated NAT bridge
claudenet, IPv6 off. Every rule below is IPv4-only, so IPv6 would be an uncovered path — off is part of the contract, not a default. claude-isolateACL — drops all egress to private space (RFC1918, CGNAT, link-local), with a single carve-out to the gateway so DNS works.- Sibling isolation, at L2 — two boxes on one bridge are switched,
never routed, so no L3 rule can separate them (learned the hard way; see
below).
security.port_isolationon every box NIC plus an nft bridge-family drop mean box A cannot exchange frames with box B at all. - No name-level reconnaissance —
dns.mode=nonestops the gateway resolving sibling names, and the bridge's resolver is pinned to public upstreams (no-resolv), so tailnet names and split-DNS zones from a host-level VPN don't resolve inside a box either. - Host firewall — instance → host is dropped except DNS/DHCP, including
the host's public IPs. Entry is
incus execover the local socket only — no inbound path exists.
The VM is the trust boundary: Claude can run arbitrary code inside and touch nothing you care about.
Measured, not claimed
Every clause above is probed live by an end-to-end drill, because the one time this contract was reasoned about instead of measured, the reasoning was wrong: box→box traffic was "covered" by an L3 drop that L2-switched frames never meet — a hole found by probing, not by reading the rules. On a bare host the drill installs the whole stack, mints a box cold, snapshots and clones it, probes every boundary from inside the boxes, and removes what it minted — currently 47 checks, 47 passing. drill/RUNS.md is the full history, including every trap that fooled a run into a wrong verdict.
bash drill/doctor.sh # read-only: is this host healthy and the stack live?
bash drill/drill.sh # FULL end-to-end — mutates the host; use a machine you own
The doctor reads ground truth, not config claims — the kernel's isolated on
flag per bridge port, the process table, the resolver actually in use — and
diagnoses the host faults that have actually happened: a wedged Incus daemon,
a dnsmasq that silently isn't serving, a VPN resolver that boxes would
inherit.
Recipes: the .claudebox/ convention
A repo that wants to be easy to stand up in a box ships an optional .claudebox/
folder — a runbook Claude reads and follows (install deps, start services,
template env, seed data, smoke-test). It is agent-facing documentation, not a
host-executed script. See docs/claudebox-recipe.md.
Uninstall
~/.local/share/claudebox/host/teardown-host.sh # boxes, network, ACL, profile, firewall
~/.local/share/claudebox/host/teardown-host.sh --purge-incus # ...and Incus itself
rm -rf ~/.local/share/claudebox ~/.local/bin/claudebox # the CLI
Non-goals
- No unattended/CI bring-up. The flow is interactive (log in, clone, ask Claude). Reproducible-by-construction provisioning is out of scope.
- No credential storage or injection by the tool. Boxes are creds-free; snapshots are the reuse mechanism, not a secrets store.