claude-hdb dbefd66dd5 fix: boxes could reach each other — isolate them at the bridge

A live probe (drill run 10) found box A's SYN arriving at box B, and B
answering with a RST. Boxes were not isolated from each other at all,
and the README's contract — "a box reaches the public internet and
nothing else" — was false.

The ACL was not wrong; it simply never saw the traffic. Two boxes on one
bridge share an L2 segment, so their frames are SWITCHED between bridge
ports and never traverse the netfilter path where an L3 rule lives. The
drop on 10.0.0.0/8 (which contains claudenet) and the default ingress
drop both looked airtight and neither ever fired. This is why the
original reasoning — "belt and braces" — was plausible and wrong.

The bridge family does see it. Its forward hook fires exactly when a
frame passes from one bridge port to another, which on claudenet means
box→box and nothing else: frames for the gateway are delivered locally,
and so is anything routed out to the internet. Dropping every forwarded
frame on the bridge isolates the boxes and costs them nothing — DHCP and
ARP are unaffected, being broadcast and delivered on INPUT.

Also: dns.mode=none, so a box can no longer ENUMERATE its siblings
through the gateway's dnsmasq. Blocked connections with open
reconnaissance is not isolation.

security.ipv4_filtering is deliberately NOT used: it breaks the box's
networking (dockerd comes up but cannot pull or run a container).

The drill now ASSERTS all of this in phase C against the real stack;
phase D's rehearsal is retired, its findings recorded.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

2026-07-14 01:29:33 +00:00

6.5 KiB

Raw Blame History

claudebox design

claudebox is a CLI that mints and manages trust-less, network-isolated VMs with Claude Code installed. It is infrastructure, not a project provisioner.

See issue #3 for the full reframe and rationale. This doc captures the durable design decisions.

Principle: separate the tool from the agent

The tool mints isolated boxes with Claude installed but unauthenticated. It knows nothing about projects, secrets, recipes, or memory.
The agent (Claude Code, inside the box) reads an optional .claudebox/ runbook in a cloned repo and acts on it. The recipe's consumer is the reasoning agent, not host machinery.

Boxes are strictly creds-free

claudebox new --name <n> launches a blank box: everything installed, no git credentials and no Claude credentials. The operator authenticates interactively inside the box:

Claude — claude → /login (paste-a-code OAuth: copy the URL, open it in your own browser, paste the code back). Works because the box is outbound-only; the tool never handles a token.
Git — the operator adds their own PAT / gh auth login inside the box.

The tool stores and injects no credentials, ever. This dissolves the multi-user problem: nothing shared, nothing committed.

Snapshots are the reuse mechanism

Re-authing every fresh box would be toil, so authenticated state is reused via snapshots, not a secrets store:

claudebox snapshot <n> [label] — checkpoint after login + clone.
claudebox new --name <n2> --from <src>[/<snapshot>] — clone an existing box or snapshot (authed state and all). Isolation is preserved: the clone keeps the claude-dev profile + claudenet + ACL.
claudebox restore <n> <snapshot> — roll a box back to a checkpoint.

The box announces itself to the agent

cloud-init installs a global ~/.claude/CLAUDE.md in every box telling Claude it is running in a claudebox (trust-less, ephemeral, creds-free) and to treat a repo's .claudebox/ folder as its bootstrap runbook. No "tell it" step, no host execution.

`.claudebox/` is optional, agent-facing documentation

Not host-executed shell. A repo that wants to be easy to stand up in a sandbox ships a runbook (prose + optional scripts the agent may run). A repo that does not, you set up by hand. The tool enforces no contract; there is no install.

What claudebox owns, and what it doesn't

Boxes are ordinary Incus instances, tagged user.claudebox=1. That makes every Incus verb a candidate feature request — rename, info, file push, on forever — and wrapping them one at a time grows a worse incus. The rule:

claudebox owns a command when it must enforce an invariant Incus cannot see: the user.claudebox=1 boundary (never touch an instance we didn't mint), the isolation stack (claude-dev profile + claudenet + ACL), or the creds-free snapshot→clone workflow. Everything else is Incus's job.

The rule cuts both ways, and that's the point:

rename is ours — not because it adds logic to incus rename, but because resolving the name is the logic: check the tag, apply --remote, and notice the box is running (Incus won't rename a running instance) so we can say "stop it first" rather than leak an Incus error.
incus config set security.nesting=false is not ours. It dismantles the trust boundary; wrapping it would imply we bless it.

Two mechanisms keep this honest.

The command table (CMDS in bin/claudebox) is the single source of truth for what exists, its synopsis, its help line, its preconditions and what runs. Dispatch and help are both rendered from it, so the help cannot describe a command that doesn't exist — the failure that produced #8. A thin verb is one row; a verb that can't be expressed as a row and enforces no invariant of ours doesn't belong in the tool.

The escape hatch — claudebox incus <box> -- <args...> — resolves and tag-checks the box, then hands the rest to Incus verbatim. It means "no" to a proxy request is not "you can't do that", and it keeps the one rail that matters: you cannot aim it at an instance claudebox didn't mint. If the command can move the box off the isolation stack (profile, network, device, security.*), it warns and proceeds — from there the trust boundary is yours to keep.

Isolation

Dedicated NAT bridge claudenet + Incus claude-isolate ACL dropping all RFC1918/CGNAT/link-local egress, plus host-firewall rules blocking instance → host. Entry is incus exec over the local socket — no inbound path. The VM is the trust boundary.

A box reaches the public internet and nothing else — including no other box. That last clause is the one that was assumed and turned out to be false, so it is spelled out here with the mechanism, and drill/ tests it on every run.

Box → host, LAN, RFC1918, CGNAT, link-local: the claude-isolate ACL.
Box → box: an nftables bridge-family rule (host/claudebox-firewall.sh). It cannot be an ACL rule. Two boxes on one bridge share an L2 segment, so their frames are switched between bridge ports and never traverse the netfilter path an L3 ACL lives on — the ACL looked airtight (it drops 10.0.0.0/8, which contains claudenet) while box→box was in fact wide open. A live probe found box A's SYN arriving at box B. The bridge family's forward hook fires exactly on port-to-port frames, which on this bridge means box→box and nothing else: gateway traffic and routed egress are delivered locally, not forwarded. Dropping every forwarded frame therefore isolates the boxes and costs them nothing.
Box → box by NAME: dns.mode=none. dnsmasq on the gateway held a record for every instance, so a box could enumerate its siblings even where it could not reach them. Blocked connections with open reconnaissance is not isolation.
IPv6: off (ipv6.address=none), and that is a contract, not a default — every rule above is IPv4-only, so IPv6 would be an uncovered path.
security.ipv4_filtering: deliberately NOT used. It breaks the box's networking (in-box Docker cannot pull or run a container). Tested, vetoed.

The rule that keeps this honest: isolation claims are tested, never reasoned about. The box→box hole existed because a plausible code reading said it could not. See drill/RUNS.md.

Non-goals

No unattended/CI bring-up — the flow is interactive.
No credential storage or injection by the tool.

6.5 KiB Raw Blame History