sbclaude

sbclaude runs Claude Code inside a throwaway Docker container with the bash sandbox and permission prompts disabled, against a configurable set of host bind mounts. The container stores nothing of its own (--rm); everything lives on the host. You only ever invoke sbclaude — it manages its own image and containers via the Docker SDK.

The host must be Linux. This is not a packaging gap that could be closed later: the host copy of claude is bind-mounted into a Linux container and executed there, so the host must supply an ELF build of it (a macOS host has a Mach-O one, which cannot run in the container); the identity mirroring reads os.getuid(); and every device and socket the run flags pass through (/dev/nvidia*, /dev/dri, /dev/kvm, /dev/bus/usb, /tmp/.X11-unix, /run/user/<uid>, /var/run/usbmuxd) is a Linux one. sbclaude refuses to start anywhere else rather than failing later with a Docker error.

There is a single image, sbclaude, built on demand. It bundles everyday coding tools (Debian slim + git, ripgrep, Node 24/Yarn, a C toolchain, gh, glab, uv, jq), formatters and linters kept at their latest upstream release (clang-format, jsonnet, jsonnetfmt, shellcheck), Qt 6 development (qt6-base-dev plus ninja), webshot (a GPU Chrome driver for screenshots and visual diffs, with Chromium baked in), and a mobile reverse-engineering toolchain (a JDK, frida, mitmproxy, dex2jar, baksmali/smali, and launchers for the host-mounted Ghidra, Android SDK, jadx, and apktool).

How it works

• The claude executable is not installed in the image. It is a self-contained native binary (Node is bundled in), so the host copy is bind-mounted read-only at run time — the image always tracks whatever version the host has. (Only glibc ≥ 2.17 is needed, which is why the base is Debian, not Alpine/musl.)
• Your identity is mirrored: UID/GID/USER/HOME are passed in and an entrypoint recreates that user inside the container. This means:
  • mounted ~/.claude files (incl. the 0600 .credentials.json) are owned correctly;
  • paths resolve identically — a host path you paste into a prompt (/home/you/dev/foo) is bind-mounted at that same absolute path inside the box.
• --dangerously-skip-permissions is passed (hence the non-root user — that flag refuses root), and a patched copy of your settings.json is mounted over the in-container one with sandbox.enabled=false, skipDangerousModePermissionPrompt=true (no bypass dialog) and tui="fullscreen". Your real settings file is never modified.
• The image auto-builds on first use and rebuilds automatically when the packaged Dockerfile/entrypoint change (tracked via a content-hash label).

Running from a non-Linux host

This is unsupported, but the reason is narrow enough to work around, and there is one trap that is not obvious until it bites.

The remaining constraint is not the operating system as such: docker resolves bind-mount sources on the daemon's filesystem, not the client's. sbclaude leans on that: ~/.claude and the project are mounted at their own absolute paths, so a path you paste into a prompt resolves identically inside the box, and the host's claude binary is mounted at /usr/local/bin/claude. Every one of those sources has to exist on the daemon. On a Linux host the client and the daemon share one filesystem and this comes free; anywhere else it is something you have to arrange.

Windows — use WSL2, which is a Linux environment rather than an emulation layer, so nothing needs arranging. Install sbclaude and claude inside the distribution, turn on Docker Desktop's WSL integration (or run dockerd in the distribution itself), and keep projects on the WSL2 filesystem rather than under /mnt/c. Client and daemon then share a filesystem exactly as they do on a native Linux host.

macOS — there is no WSL2 equivalent, so either run sbclaude inside a Linux VM that also runs the daemon, or log in to a Linux machine and run it there. Pointing DOCKER_HOST at a remote daemon from the Mac is not a third option: sbclaude refuses to start there, and none of the paths it would name exist on that daemon anyway.

In a VM, the device flags (--gpu, --x11, --wayland, --android, --usb, --ios) want host hardware it does not have, and --net host means the VM's network, so an MCP server on the outer machine's localhost is out of reach.

The trap — a bind mount whose source is missing on the daemon is not an error; docker creates an empty directory and mounts that instead. Usually this fails loudly, because an empty directory at /usr/local/bin/claude cannot be executed, so the box exits 126 the moment the entrypoint reaches it and the reason lands in syslog (see When a box does not start). The patched settings.json is the exception. It is written with tempfile, so it lives in the client's /tmp; where that is not also the daemon's, an empty directory lands over ~/.claude/settings.json and the box comes up without skipDangerousModePermissionPrompt=true or tui="fullscreen", so it stops at the bypass dialog those settings exist to remove. The bash sandbox stays off either way — the image's managed settings force that, and nothing mounted can override them. Set TMPDIR to a directory both sides can see.

Install

uv tool install .         # or: pipx install .
# the image builds itself on first `sbclaude run`; or pre-build:
sbclaude build

Usage

sbclaude                          # run claude; cwd is the project (writable)
sbclaude run -p ~/dev/foo         # explicit project dir (writable, becomes workdir)
sbclaude run -r /data -w ~/scratch   # extra read-only / read-write mounts
sbclaude run --re --x11           # Ghidra/Android mounts + GUI passthrough
sbclaude run -- --version         # everything after -- goes to claude
sbclaude ls                       # list running sbclaude containers
sbclaude stop [--all]             # stop this project's boxes (or all with --all)
sbclaude shell                    # debug shell in this project's box, as your user
sbclaude shell --root             # ... as root instead
sbclaude build [--no-cache]       # (re)build the image
sbclaude delete-image             # remove the sbclaude image
sbclaude config                   # show the config file path
sbclaude scaffold-noclip [DIR]    # write a noclip visual-regression harness into a project

webshot

Inside the box, webshot drives a real GPU-accelerated Chromium (baked into the image, so no project needs its own browser download), screenshots it, and compares the result against a baseline. It is deliberately generic: it knows about browsers, waiting, settling, screenshots, and image comparison, and nothing about any particular application — per-project knowledge is supplied with --wait-fn/--eval.

webshot doctor                                          # report the renderer actually in use
webshot shot http://localhost:3000 --out out.png --wait '#app'
webshot check http://localhost:3000 --baseline base.png
webshot compare a.png b.png --diff d.png

Exit codes: 0 ok, 1 regression, 2 software renderer (i.e. the GPU is not really being used), 3 missing baseline, 64 usage. Pair it with --gpu, and with --wayland when a headed window is wanted.

scaffold-noclip

Writes the project-side glue for a visual-regression workflow built on the webshot tool that ships in the image, into DIR (default: the current directory):

viewer-tests/nc.mjs        runner -- talks only to the app's own API, copied verbatim per project
viewer-tests/views.json    the only file you edit: which scene, and where the camera goes
viewer-tests/baselines/    approved output; effectively source, nothing can regenerate it
viewer-tests/out/          captures and diffs, rewritten every run
NEW_GAME.md                getting-started guide, including the non-obvious failure modes

sbclaude scaffold-noclip --scene MyGame/Level1   # substitute the scene id while writing
sbclaude scaffold-noclip ~/dev/foo --force       # overwrite existing files

Existing files are never overwritten without --force: an edited views.json or a blessed baseline is judgement that cannot be regenerated.

You can run several boxes against the same project directory at once. Each run gets a unique container name — the project name plus a short random suffix — so there is no name collision; override it with -n. sbclaude ls lists them all, sbclaude stop stops every box for the current project, and sbclaude shell attaches to it (or asks you to pass -n NAME when more than one is running).

run flags

When a box does not start

A box that never gets going — docker exits 125–127, or the container dies within seconds — is launched again, up to three attempts, then sbclaude exits with the last code. A session that ran and then ended is never retried, whatever its exit code. Every attempt is logged to syslog with docker's own message and the docker run command behind it; read them with journalctl -t sbclaude. The terminal cannot be trusted here, because the fullscreen TUI erases its own screen on exit; --no-fullscreen keeps a failing session's output on the normal screen.

--gpu probes for a GPU with nvidia-smi -L before asking Docker for one, retrying a few times because that probe is also what wakes a card the kernel has parked. If none answers, the box starts without --gpus all (the DRM render nodes are still passed): an unresolvable nvidia.com/gpu=all is a warning to the daemon but death to the container.

Configuration

All options live under a [tool.sbclaude] table. They are read from the global file at ~/.config/sbclaude/config.toml (path from platformdirs) and, for run, overlaid with the target project's pyproject.toml [tool.sbclaude] table — project values win, so a repo can pin its own defaults. All keys optional:

[tool.sbclaude]
gpg = true       # mount the GnuPG home + agent for signing
network = "host" # default; "bridge" to isolate the box's network
re = true        # enable the Ghidra + Android mounts together
ssh = true       # mount ~/.ssh read-only + forward the ssh-agent for SSH git remotes
sudo = true      # passwordless sudo in the box (drops no-new-privileges)
wayland = true   # forward the Wayland socket for GUI apps
x11 = true       # forward X11 for GUI apps
# image = "custom:latest"         # force a different image
# debian_mirror = "http://ftp.us.debian.org/debian" # apt mirror for image builds
# memory = "8g"                   # override the auto host-RAM cap ("0" disables)
# cpus = "4"                      # cap CPUs (uncapped by default)
# recover = true                  # install cc-session-recover into every project (off by default)
# modify = false                  # never write into the project directory (same as --no-modify)
# venv_dir = "/venv-cache"        # hold the box's virtualenv here (pair with a docker_args volume)
# fullscreen = false              # do not force the fullscreen TUI (keeps start-up errors visible)
pass_env = ["AWS_REGION"]                          # forward host vars (AWS_PROFILE is default)
ro = ["~/dev*", "~/ghidra_scripts", "~/Downloads"] # read-only mounts (globs + ~ ok)
rw = []                                            # the project dir is always rw automatically

[tool.sbclaude.env] # inject fixed vars (e.g. Amazon Bedrock)
CLAUDE_CODE_USE_BEDROCK = "1"

The toggle keys re, ghidra, android, gpu, usb, ios, wayland, x11, ssh, gpg, and sudo mirror the matching run flags and default to false; setting one is the same as always passing that flag.

Debian mirror — when deb.debian.org is slow, point image builds at a faster archive mirror with debian_mirror (or --debian-mirror on build/run). Only the image's main archive URI is rewritten; debian-security keeps pointing at deb.debian.org.

ro/rw accept globs (~/dev* → every matching dir) and ~; non-matching or missing paths are dropped. Every path is mounted at its real absolute path, so a host path you paste into a prompt resolves inside the box. The project (cwd or -p) is always read-write and overlays any read-only parent (e.g. ~/dev ro + ~/dev/proj rw → proj is writable).

Environment variables — inject with the [tool.sbclaude.env] table, forward host values by name with pass_env, or per-run with -e KEY=VALUE (precedence: managed defaults < [tool.sbclaude.env] < pass_env < -e). AWS_PROFILE is forwarded by default. This is how you point the box at a different backend such as Amazon Bedrock (CLAUDE_CODE_USE_BEDROCK=1 + your AWS_* vars; mount ~/.aws via ro if you use profiles).

uv project environment — by default the box sets UV_PROJECT_ENVIRONMENT to .sbclaude-venv, so uv sync/uv run build the virtualenv there instead of writing .venv into your bind-mounted project. A host-built .venv hard-codes the host interpreter path and would not resolve in the box anyway, so the two are kept apart; the host's .venv is additionally re-mounted read-only so nothing in the box can corrupt it. Beside the project rather than inside the container because the container is --rm, so a virtualenv in it would be rebuilt from nothing every run. /.sbclaude-venv/ is appended to the project's .gitignore (once, and only when it is not already there) so it stays out of git status. Override the path with venv_dir (below) or by setting your own UV_PROJECT_ENVIRONMENT (via [env] or -e), or disable the behaviour entirely with manage_uv_env = false.

None of it happens unless the project is a Python one: a pyproject.toml, setup.py, setup.cfg, requirements.txt, or Pipfile at the top, or a .py file anywhere below it (ignoring dotted directories, node_modules, vendor, and target). A project with no Python in it gets no .sbclaude-venv, no .gitignore line, and no virtualenv variables at all.

There is no clean equivalent for Node — node_modules is fixed to the package root by Node's resolver (only Yarn Berry's YARN_NODE_LINKER=pnp removes it, at the cost of changing module resolution), and node_modules built on the box's Linux is usually reusable on a Linux host anyway, so it is left untouched.

A virtualenv that outlives the box — --venv-dir DIR (config key venv_dir) puts the box's virtualenv in DIR instead of beside the project. Combined with a volume in docker_args, the environment survives the box without anything being written into the project:

[tool.sbclaude]
docker_args = ["-v", "sbclaude-venv-cache:/venv-cache"]
venv_dir = "/venv-cache"

DIR gets one subdirectory per project, named after it with a digest of its absolute path appended (/venv-cache/sbclaude-venv-foo-1a2b3c), so a single volume serves every project and ~/dev/foo and ~/work/foo do not end up sharing an environment. The directory is created and handed to your user by the entrypoint, which is what makes a fresh (root-owned) named volume usable. venv_dir applies even with manage_uv_env = false — naming a directory is asking for the environment to be managed — but not to a project with no Python in it.

Leaving the project untouched — --no-modify (or modify = false) stops sbclaude itself writing anything into the directory it was launched from. Three things are suppressed:

• the /.sbclaude-venv/ line is not appended to .gitignore;
• UV_PROJECT_ENVIRONMENT points inside the container instead of beside the project, so neither sbclaude nor a later uv sync creates a virtualenv there. That path is under /tmp and dies with the box, unless venv_dir names somewhere that persists;
• the start-up uv sync is skipped (it refreshes uv.lock), as is the cc-session-recover install, which writes into .claude/ and HANDOFF.md. Asking for both --session-recover and --no-modify prints a note and leaves recovery off.

This constrains sbclaude, not Claude: the project is still bind-mounted read-write and the agent can edit it exactly as before.

Alternate config dir — if CLAUDE_CONFIG_DIR is set on the host, sbclaude mounts that directory (its .claude.json, settings.json, history) and points claude at it inside the box instead of ~/.claude.

Session recovery

cc-session-recover lets Claude Code pick a long-running task back up after a quota or rate-limit pause. It is off by default — enable it per run with --session-recover, or for every run with recover = true in the config. When enabled, the container entrypoint runs the tool's install-into-project.sh against the project before launching claude, which sets up its SessionStart and Stop hooks and a HANDOFF.md.

The tool is vendored into the image as a git clone (not the npm package) at /opt/cc-session-recover, with PR #2 (safer watcher argument handling and the opt-in CC_REMIND_MODE prompt-injection limit) applied on top. Two further patches keep the installer tidy: it no longer copies settings.example.json into the project, and its own .gitignore handling is disabled. The installer writes into the project's .claude/ (hooks and settings.local.json, merged with jq) and a HANDOFF.md; because the project is bind-mounted read-write, those files land in your real repository and persist, which is why this is opt-in. Afterwards the entrypoint appends the recovery artifacts (HANDOFF.md, auto-continue.md, session-recover.js, session-recover.yaml, standing-instructions.md, statusline-quota-cache.sh, and the three hook scripts) to the project's .gitignore, each only when absent. If the install fails, the box aborts rather than starting a session that silently lacks recovery.

MCP servers

MCP server configs live in your mounted ~/.claude.json, so they carry into the box — but the server command must be runnable inside the container. A host Python venv won't work: its bin/python symlinks to a host-only interpreter (e.g. /usr/bin/python3.13), which doesn't exist in the box → ENOENT. The image ships Python 3 in a virtualenv at /opt/venv (with the mcp SDK and pre-commit pre-installed) and uv, so point the command at one of:

• uv run /abs/path/to/server.py — best: reads the script's PEP 723 deps, works on the host too. Example: claude mcp add ghidra -- uv run ~/dev/ghidra-mcp/bridge_mcp_ghidra.py.
• /opt/venv/bin/python3 /abs/path/to/server.py — the container's venv Python (has mcp pre-installed); container-only. /opt/venv/bin is first on PATH, so a bare python3 resolves here too.

A server that talks to a process on the host (e.g. a Ghidra GUI on localhost) works out of the box because the box defaults to host networking; pass --net bridge only if you want to isolate it.

Git over SSH and commit signing

Pushing over SSH and signing commits need the host's private keys, which are not mounted by default (the box is a fully-autonomous, no-prompt agent — see Hardening). Opt in per run, or globally with ssh = true and gpg = true under [tool.sbclaude]:

• --ssh bind-mounts ~/.ssh read-only, so SSH remotes authenticate with the host's keys and known_hosts. Read-only means newly-learnt host keys are not written back. It also forwards the host's ssh-agent: $SSH_AUTH_SOCK is bind-mounted at its own path and set in the box, so keys that are passphrase-protected or held in a hardware token still work — the host agent does the signing and the secret never enters the container. Without this a box holding only encrypted key files would stall on a passphrase prompt nothing can answer. Symlinks inside ~/.ssh are followed too: a config (or key) linked into a dotfiles repository has its target mounted read-only at the same path, so the link does not dangle and every Host alias keeps working.
• --gpg bind-mounts the host GnuPG home (read-write — gpg needs to write lock files and the trustdb) and overlays the host's live gpg-agent socket at both of the places the box's gpg may look for it: ~/.gnupg/S.gpg-agent and /run/user/<uid>/gnupg/S.gpg-agent. Which one applies depends on whether /run/user/<uid> exists in the box, and --wayland and --ssh both make it exist, so both are mounted rather than guessed. The host agent performs the signing and owns the secret keys, so a cached passphrase carries over and any pinentry prompt appears on the host. The image ships gnupg and openssh-client; your mounted ~/.gitconfig (with user.signingkey/commit.gpgsign) does the rest.

sbclaude run --ssh --gpg -p ~/dev/foo   # inside: git push, git commit -S both work

Your ~/.gitconfig is always mounted read-only, along with every file it pulls in via an [include] path = ... directive (resolved with git config --includes), so a split config carries into the box intact.

Root inside the box (--sudo)

The box runs as your mirrored user, and so does sbclaude shell — it opens a login shell as that user, in the project directory. sbclaude shell --root gives a root shell instead; that one asks the Docker daemon to start the process as root, so it works on any box, escalating nothing inside it.

sudo is a different matter, because it is what lets the agent become root — to apt-get install a missing library, say. It is off by default and enabled with --sudo (or sudo = true). The box then gets a NOPASSWD: ALL sudoers drop-in for your user, so sudo su and sudo <anything> work without a password:

sbclaude run --sudo -p ~/dev/foo   # inside: sudo su, sudo apt-get install ... both work

The catch, and why this is a flag rather than the default: sudo is setuid-root, and the no_new_privs kernel flag that --security-opt no-new-privileges sets makes the kernel ignore the setuid bit on every execve — sudo detects this and refuses to run. The two cannot coexist, so --sudo drops no-new-privileges from the hardening set. Everything else stays: the capability set is still ALL dropped plus the same six, so container root here has no CAP_SYS_ADMIN, no CAP_MKNOD, and no CAP_NET_ADMIN. What it does gain is CAP_DAC_OVERRIDE over every mounted path, i.e. the file permissions on your read-write mounts stop being a boundary. Read-only mounts stay read-only — that is enforced by the kernel, not by permissions.

The image ships sudo but still strips every setuid bit at build time; the entrypoint restores it on sudo alone, and only for a box started with --sudo. A box started without it therefore contains no setuid binary at all.

RE toolchain (--re)

Mounted from the host (your exact versions): Ghidra (analyzeHeadless, ghidraRun), Android SDK (adb, emulator, sdkmanager, avdmanager), jadx, apktool, plus ~/.android (adb keys + AVDs) and /dev/kvm for emulator acceleration.

Installed in the image: Temurin JDK 21 (Ghidra 12 needs it), build-essential + binutils, frida + frida-tools (pinned to the host version), mitmproxy, dex2jar, baksmali/smali, CLI audio tools (ffmpeg, sox, flac, vorbis-tools, opus-tools, lame, mpg123, wavpack, shntool, plus vgmstream-cli for game audio), CLI image tools (ImageMagick, zbar, deark, pngdefry for -iphone PNGs), czkawka-cli (duplicate/similar finder), and the X11/GL/audio libs the mounted GUI binaries need.

sbclaude run --re --x11 -p ~/dev/some-apk-re
#   inside: jadx -d work/jadx-out base/classes*.dex
#           apktool d base -o work/axml-decoded
#           analyzeHeadless ~/dev/x-re proj -import lib/arm64-v8a/foo.so
#           adb devices ; emulator -avd ford-x86_64-api35 -writable-system &
#           frida -U -f com.x.y -l hook.js

--x11 mounts /tmp/.X11-unix + your session xauth cookie and sets DISPLAY/ XAUTHORITY. Java/Swing (Ghidra) renders through XWayland (:0). If a GUI fails with a cookie error, run on the host: xhost +SI:localuser:$USER.

--wayland is the better option when the app speaks Wayland: it forwards the compositor socket alone, and a Wayland client cannot read other windows or inject input into them, whereas an X11 cookie grants exactly that over the whole session. The socket is re-homed under the box's own /run/user/<uid> and WAYLAND_DISPLAY/XDG_RUNTIME_DIR are set to match.

--gpu uses the NVIDIA container runtime when it is present and also passes through the DRM render nodes (/dev/dri/renderD*), so Mesa on AMD/Intel and Vulkan/VA-API work too. The entrypoint re-adds the device groups Docker granted, and synthesises the glvnd/Vulkan/GBM manifests the NVIDIA toolkit does not install, so GL, EGL, and Vulkan clients get the real driver instead of a software fallback.

--ios targets an iOS device attached to the host. The host runs usbmuxd (it owns the USB device), and frida's usbmux backend reaches the device through that daemon's socket — so instead of claiming raw USB (which would clash with the host usbmuxd), --ios bind-mounts /var/run/usbmuxd plus the host's /var/lib/lockdown pairing records. Combine with --re so frida is present:

sbclaude run --re --ios -p ~/dev/some-ios-re
#   inside: frida-ls-devices            # the host's device shows up over usbmux
#           frida -U -f com.x.y -l hook.js

The host needs usbmuxd running and the device already paired (trusted). If frida sees no device, confirm idevice_id -l works on the host first.

Hardening

This deliberately removes Claude's own sandbox and permission prompts, so the box leans on Docker for confinement instead. Every run is hardened by default (defense-in-depth — it limits blast radius, it is not a guarantee):

• Build time: minimal Debian slim, --no-install-recommends + cleaned apt lists, no secrets baked in (the claude binary and all auth are bind-mounted), OCI provenance labels, and all setuid/setgid bits stripped from the image.
• Run time: --security-opt no-new-privileges (unless --sudo asked for the opposite — see Root inside the box), --cap-drop ALL plus only the six caps the root entrypoint needs to create the mapped user, drop to it via gosu, and let tini (PID 1) forward signals such as SIGWINCH to the non-root child (CHOWN, DAC_OVERRIDE, FOWNER, KILL, SETUID, SETGID), a --pids-limit, a non-root mapped user, and the default seccomp/AppArmor profiles (never disabled). The claude process itself ends up with an empty effective capability set.
• Can't lock the host: --pids-limit stops fork bombs, and --memory with an equal --memory-swap bounds RAM with no extra swap, so a runaway box is OOM-killed instead of thrashing the host into a freeze. The default cap is derived from host RAM (reserving the larger of 2 GiB or an eighth for the host); set memory to override or "0" to disable, and cpus to cap CPU (uncapped by default — saturation slows but does not lock). These are enforced by cgroups, i.e. the systemd cgroup driver on a systemd host — a separate systemd-run wrapper is unnecessary (and would not bound the container, which runs under the Docker daemon's cgroup, not the CLI's).

Disable per run with --no-harden, globally with harden = false in config (this also drops the resource caps), and add your own Docker flags (e.g. --read-only) via docker_args = [...].

Still: the containerized Claude can run any command and read/write every mounted path without asking, with unrestricted network. Keep writable mounts minimal (the config defaults to read-only for everything but the project) and don't mount secrets you don't want a fully-autonomous agent to touch. This is why --ssh, --gpg, and --sudo are opt-in: the first two expose your private SSH and GPG key material (the GnuPG home read-write) to that agent, and the third hands it container root. Forwarding an agent socket does not hand over the secret itself, but it does let the box ask the host agent to sign with any key it holds, for as long as the box runs.