Sandbox

This document explains tool execution confinement on Linux and the existing whole-CLI sandbox methods.

Linux tool execution sandbox

On Linux, install bwrap (Bubblewrap) at /usr/bin/bwrap (the supported system path) for the strongest supported boundary, then enable the tool execution sandbox in your User settings (~/.qwen/settings.json, or the directory selected by QWEN_HOME) or administrator System settings:

{
  "tools": {
    "executionSandbox": {
      "backend": "auto",
      "filesystem": "workspace-write",
      "network": "closed"
    }
  }
}

filesystem and network are required. backend defaults to auto, which tries bwrap first. When bwrap is unavailable and network is open, auto can use the bundled Landlock helper on kernels with Landlock ABI 3 or newer. Set backend to bwrap or landlock to require that exact implementation. Landlock cannot enforce network: closed or hide explicit masked paths. Either requirement rejects Landlock before any command starts; automatic fallback never discards a remaining path mask. Use read-only to deny workspace writes, or workspace-write to permit writes inside the current canonical workspace. Each command also gets private scratch space. Ordinary includeDirectories settings do not grant write access. Keep the workspace separate from the Qwen installation, user configuration and runtime state directories.

The CLI, model transport, authentication and session storage remain on the host. Shell, terminal !, prompt shell interpolation, Monitor, Read/Write/Edit and supported nested in-process Agent/Code Mode calls use the same runtime policy. With bwrap, network: closed blocks ordinary host/external IP connections from commands, while open shares the host network. Reads remain broad: this mode does not hide host secrets or promise to isolate pathname Unix sockets and host services. Approval and YOLO do not expand the filesystem or network policy. Standard streams are explicit caller-provided capabilities: an already-connected socket, TTY or character device remains usable even with network: closed. Closed networking does not revoke those existing descriptors; avoid supplying a host-service connection as stdin when that access is unwanted. Host-backed file and named-FIFO stdin require the bundled Linux x64/arm64 input helper and are copied through a private Unix stream socket with its write half shut; regular-file offsets advance only by bytes consumed from that socket, and an idle FIFO writer does not delay command completion. Read fd 0 directly: this input is a socket rather than a FIFO, and reopening it through /dev/stdin or /proc/self/fd/0 is unsupported. SIGINT/SIGTERM cancel the command but still drain accepted stdout/stderr bytes; a closed output stops that stream with exit 141 while the other valid output drains.

Landlock reports partial enforcement because it restricts pathname reads, writes and directory mutations but does not create PID or network namespaces and cannot currently restrict every metadata operation, including chmod, chown, extended attributes and timestamps. A Landlock command can see host processes and uses host networking. Detached descendants retain the Landlock filesystem restrictions but do not get bwrap’s PID-namespace lifetime boundary. Open file descriptors retain their existing access. Qwen Code streams host-backed file and named-FIFO stdin through a private read-only socket instead of passing those descriptors to the payload; sockets, character devices and anonymous pipes can be shared. Output standard streams remain the caller’s explicit destinations. The execution-status descriptor closes on payload exec.

Landlock also has a narrower writable device surface: it grants writes to /dev/null, but does not create bwrap’s private /dev and /proc mounts. Opening /dev/full or /dev/tty for writing and creating files in /dev/shm are denied, so programs requiring these operations may fail. Use the private scratch directory exposed through TMPDIR, TMP, and TEMP for temporary files. Already-open, caller-provided standard streams keep their existing access. Startup’s Landlock policy probe requires /usr/bin/true; a host missing it fails the probe even if its kernel supports Landlock.

Workspace settings cannot enable, disable or modify this policy, even in a trusted project. Effective precedence is System over User over SystemDefaults, selecting a complete policy object. Values must be literals; environment substitutions, unknown fields and incomplete objects are rejected. --bare and --safe-mode retain operator confinement. Changes require a new runtime. A malformed or unreadable SystemDefaults, User or System settings file now blocks startup (configuration exit code 52), instead of silently resetting unknown policy to empty settings. The error names the file and repair action. Repair its JSON object or restore read access, then restart. A User .corrupted copy is only a reference; the original operator file is never cleared. Workspace-only recovery remains available without operator confinement. Malformed linked Workspace settings (including a linked .qwen directory or a hard-linked file) require manual repair; automatic recovery refuses before changing the original or a previous corruption copy. Valid linked JSON remains readable.

This initial public mode supports the ordinary headless CLI and both terminal UIs. ACP, qwen serve and web terminals explicitly reject it. MCP/LSP, executable hooks/extensions, worktree/arena management, external agents, speculative execution, file checkpoints, automatic skill/memory maintenance, and unported process or mutation tools are disabled or rejected. Ordinary same-workspace in-process agents cannot add hooks, MCP servers, external executors or worktree isolation. Custom status commands, external Markdown/image renderers, IDE detection and connection, file restore, and host Git diff previews are disabled. Conversation-only rewind remains available.

Explicit operator diagnostics such as /doctor, opening or editing files and folders through /memory, and changing settings through /skills remain trusted host actions. These management commands are not exposed as model-invocable commands. Picking a skill fills the input field; submitting /skill-name currently reports an unsupported-mode error before preparing arguments or updating usage records. Skill command execution and the model Skill tool remain unavailable until their preparation writes are confined. Automatic memory and skill maintenance remain disabled in this runtime even if you change their saved settings.

Inspect and verify the active boundary before running a task:

qwen sandbox
qwen sandbox --verify
qwen sandbox -- sh -c 'printf "confined command\n"'
qwen -y -p "Update the project and run its tests"

Linked Git worktrees receive write access to the admitted workspace only. Git common-directory and worktree metadata outside that workspace stay read-only; this mode does not implicitly authorize changing another checkout’s repository metadata. Operations requiring those writes can be refused.

The report names the tool boundary, requested/effective backend, enforcement level, Landlock ABI when applicable, workspace, filesystem and command network policy. --verify checks workspace writes, denial against a file known to be writable on the host, the expected private or shared PID namespace identity, and the selected network namespace. A passed backend probe establishes admission, not authenticated full-session viability under all inherited environment settings. Run an ordinary task afterwards to test its required tools and services. If verification cannot create its host-side temporary fixture, check TMPDIR and directory permissions; this is separate from a boundary refusal. Backend setup errors fail before payload execution and never rerun the command on the host. The one-command subcommand forwards literal arguments and redirected stdin while preserving stdout and stderr byte-for-byte; use terminal ! for interactive PTY commands.

Migrating from whole-CLI bwrap

Replace --sandbox bwrap and remove tools.sandbox: "bwrap", QWEN_SANDBOX=bwrap, QWEN_SANDBOX_NET and QWEN_SANDBOX_PROXY_COMMAND, then configure tools.executionSandbox as above. Restart from a shell outside the old sandbox; inherited SANDBOX=bwrap is rejected. proxied is not a supported network policy. Do not combine this mode with Docker, Podman, Seatbelt or an inherited whole-CLI sandbox marker. Old bwrap settings produce an explicit migration error.

The Docker, Podman and macOS Seatbelt methods below still run the whole CLI in their existing environment.

Prerequisites

Before using sandboxing, you need to install and set up Qwen Code:

npm install -g @qwen-code/qwen-code

To verify the installation

qwen --version

Overview of sandboxing

Sandboxing isolates potentially dangerous operations (such as shell commands or file modifications) from your host system, providing a security barrier between the CLI and your environment.

The benefits of sandboxing include:

• Security: Prevent accidental system damage or data loss.
• Isolation: Limit file system access to project directory.
• Consistency: Ensure reproducible environments across different systems.
• Safety: Reduce risk when working with untrusted code or experimental commands.

Sandboxing methods

Your ideal method of sandboxing may differ depending on your platform and your preferred container solution.

1. macOS Seatbelt (macOS only)

Lightweight, built-in sandboxing using sandbox-exec.

Default profile: permissive-open - restricts writes outside the project directory, but allows most other operations and outbound network access.

Best for: Fast, no Docker required, strong guardrails for file writes.

2. Container-based (Docker/Podman)

Cross-platform sandboxing with complete process isolation.

By default, Qwen Code uses a published sandbox image (configured in the CLI package) and will pull it as needed.

The container sandbox mounts your workspace and your ~/.qwen directory into the container so auth and settings persist between runs.

Best for: Strong isolation on any OS, consistent tooling inside a known image.

Choosing a method

• On macOS:
  • Use Seatbelt when you want lightweight sandboxing (recommended for most users).
  • Use Docker/Podman when you need a full Linux userland (e.g., tools that require Linux binaries).
• On Linux/Windows:
  • Use Docker or Podman.

Quickstart

# Enable sandboxing with command flag
qwen -s -p "analyze the code structure"
 
# Or enable sandboxing for your shell session (recommended for CI / scripts)
export QWEN_SANDBOX=true   # true auto-picks a provider (see notes below)
qwen -p "run the test suite"
 
# Configure in settings.json
{
  "tools": {
    "sandbox": true
  }
}

Configuration

Enable sandboxing (in order of precedence)

1. Environment variable: QWEN_SANDBOX=true|false|docker|podman|sandbox-exec
2. Command flag / argument: -s, --sandbox, or --sandbox=<provider>
3. Settings file: tools.sandbox in your settings.json (e.g., {"tools": {"sandbox": true}}).

--sandbox and -s accept an optional value: true, false, docker, podman or sandbox-exec. For automatic selection with a prompt, use qwen --sandbox=true "query" or qwen --sandbox -p "query". Unlike the old boolean-only parser, qwen --sandbox "query" now treats query as a backend value and rejects an unknown backend. Options before -- use the same grammar even after positional words; use -p "..." for literal option text within a prompt. -- stops option parsing, but its trailing tokens are not forwarded as the default-command prompt. Session sandbox flags are not accepted on management subcommands, whose -s aliases retain their own meaning.

Configure the sandbox image (Docker/Podman)

• CLI flag: --sandbox-image <image>
• Environment variable: QWEN_SANDBOX_IMAGE=<image>
• Settings file: tools.sandboxImage in your settings.json (e.g., {"tools": {"sandboxImage": "ghcr.io/qwenlm/qwen-code:0.14.1"}})

Priority order (highest to lowest):

1. --sandbox-image
2. QWEN_SANDBOX_IMAGE
3. tools.sandboxImage
4. Built-in default image from the CLI package (for example ghcr.io/qwenlm/qwen-code:<version>)

settings.env.QWEN_SANDBOX_IMAGE also works as a generic env injection mechanism, but tools.sandboxImage is the preferred persistent setting.

Custom images are user-managed. Rebuild them with an up-to-date Qwen Code installation to receive the safe update handoff; older images may still use their original in-process updater.

macOS Seatbelt profiles

Built-in profiles (set via SEATBELT_PROFILE env var):

• permissive-open (default): Write restrictions, network allowed
• permissive-closed: Write restrictions, no network
• permissive-proxied: Write restrictions, network via proxy
• restrictive-open: Strict restrictions, network allowed
• restrictive-closed: Maximum restrictions
• restrictive-proxied: Strict restrictions, network via proxy

Custom Seatbelt profiles (macOS)

To use a custom Seatbelt profile:

1. Create a file named .qwen/sandbox-macos-<profile_name>.sb in your project.
2. Set SEATBELT_PROFILE=<profile_name>.

Custom Sandbox Flags

For container-based sandboxing, you can inject custom flags into the docker or podman command using the SANDBOX_FLAGS environment variable. This is useful for advanced configurations, such as disabling security features for specific use cases.

Example (Podman):

To disable SELinux labeling for volume mounts, you can set the following:

export SANDBOX_FLAGS="--security-opt label=disable"

Multiple flags can be provided as a space-separated string:

export SANDBOX_FLAGS="--flag1 --flag2=value"

Network proxying (all sandbox methods)

If you want to restrict outbound network access to an allowlist, you can run a local proxy alongside the sandbox:

• Set QWEN_SANDBOX_PROXY_COMMAND=<command>
• The command must start a proxy server that listens on :::8877

This is especially useful with *-proxied Seatbelt profiles.

For a working allowlist-style proxy example, see: Example Proxy Script.

Linux UID/GID handling

On Linux, Qwen Code defaults to enabling UID/GID mapping so the sandbox runs as your user (and reuses the mounted ~/.qwen). Override with:

export SANDBOX_SET_UID_GID=true   # Force host UID/GID
export SANDBOX_SET_UID_GID=false  # Disable UID/GID mapping

Troubleshooting

Common issues

“Operation not permitted”

• Operation requires access outside sandbox.
• On macOS Seatbelt: try a more permissive SEATBELT_PROFILE.
• On Docker/Podman: verify the workspace is mounted and your command doesn’t require access outside the project directory.

Missing commands

• Container sandbox: add them via .qwen/sandbox.Dockerfile or .qwen/sandbox.bashrc.
• Seatbelt: your host binaries are used, but the sandbox may restrict access to some paths.

Java not available in Docker sandbox

The official Qwen Code Docker image is intentionally minimal to keep the image small, secure, and fast to pull. Different users require different language runtimes (Java, Python, Node.js, etc.), and bundling all environments into a single image is not practical. Therefore, Java is not included by default in the Docker sandbox.

If your workflow requires Java, you can extend the base image by creating a .qwen/sandbox.Dockerfile in your project:

FROM ghcr.io/qwenlm/qwen-code:latest
 
RUN apt-get update && \
    apt-get install -y openjdk-17-jre && \
    apt-get clean && \
    rm -rf /var/lib/apt/lists/*

Then rebuild the sandbox image:

QWEN_SANDBOX=docker BUILD_SANDBOX=1 qwen -s

For more details on customizing the sandbox, see Customizing the sandbox environment.

Network issues

• Check sandbox profile allows network.
• Verify proxy configuration.

Debug mode

DEBUG=1 qwen -s -p "debug command"

Note: If you have DEBUG=true in a project’s .env file, it won’t affect the CLI due to automatic exclusion. Use .qwen/.env files for Qwen Code-specific debug settings.

Inspect sandbox

# Check environment
qwen -s -p "run shell command: env | grep SANDBOX"
 
# List mounts
qwen -s -p "run shell command: mount | grep workspace"

Security notes

• Sandboxing reduces but doesn’t eliminate all risks.
• Use the most restrictive profile that allows your work.
• Container overhead is minimal after the first pull/build.
• GUI applications may not work in sandboxes.

Related documentation

• Configuration: Full configuration options.
• Commands: Available commands.
• Troubleshooting: General troubleshooting.