llmx/docs/sandbox.md

Sandbox & approvals

What LLMX is allowed to do is governed by a combination of sandbox modes (what LLMX is allowed to do without supervision) and approval policies (when you must confirm an action). This page explains the options, how they interact, and how the sandbox behaves on each platform.

Approval policies

LLMX starts conservatively. Until you explicitly tell it a workspace is trusted, the CLI defaults to read-only sandboxing with the read-only approval preset. LLMX can inspect files and answer questions, but every edit or command requires approval.

When you mark a workspace as trusted (for example via the onboarding prompt or /approvals → “Trust this directory”), LLMX upgrades the default preset to Auto: sandboxed writes inside the workspace with AskForApproval::OnRequest. LLMX only interrupts you when it needs to leave the workspace or rerun something outside the sandbox.

If you want maximum guardrails for a trusted repo, switch back to Read Only from the /approvals picker. If you truly need hands-off automation, use Full Access—but be deliberate, because that skips both the sandbox and approvals.

Defaults and recommendations

• Every session starts in a sandbox. Until a repo is trusted, LLMX enforces read-only access and will prompt before any write or command.
• Marking a repo as trusted switches the default preset to Auto (workspace-write + ask-for-approval on-request) so LLMX can keep iterating locally without nagging you.
• The workspace always includes the current directory plus temporary directories like /tmp. Use /status to confirm the exact writable roots.
• You can override the defaults from the command line at any time:
  • llmx --sandbox read-only --ask-for-approval on-request
  • llmx --sandbox workspace-write --ask-for-approval on-request

Can I run without ANY approvals?

Yes, you can disable all approval prompts with --ask-for-approval never. This option works with all --sandbox modes, so you still have full control over LLMX's level of autonomy. It will make its best attempt with whatever constraints you provide.

Common sandbox + approvals combinations

Intent	Flags	Effect
Safe read-only browsing	--sandbox read-only --ask-for-approval on-request	LLMX can read files and answer questions. LLMX requires approval to make edits, run commands, or access network.
Read-only non-interactive (CI)	--sandbox read-only --ask-for-approval never	Reads only; never escalates
Let it edit the repo, ask if risky	--sandbox workspace-write --ask-for-approval on-request	LLMX can read files, make edits, and run commands in the workspace. LLMX requires approval for actions outside the workspace or for network access.
Auto (preset; trusted repos)	--full-auto (equivalent to --sandbox workspace-write + --ask-for-approval on-request)	LLMX runs sandboxed commands that can write inside the workspace without prompting. Escalates only when it must leave the sandbox.
YOLO (not recommended)	--dangerously-bypass-approvals-and-sandbox (alias: --yolo)	No sandbox; no prompts

Note: In workspace-write, network is disabled by default unless enabled in config ([sandbox_workspace_write].network_access = true).

Fine-tuning in config.toml

# approval mode
approval_policy = "untrusted"
sandbox_mode    = "read-only"

# full-auto mode
approval_policy = "on-request"
sandbox_mode    = "workspace-write"

# Optional: allow network in workspace-write mode
[sandbox_workspace_write]
network_access = true

You can also save presets as profiles:

[profiles.full_auto]
approval_policy = "on-request"
sandbox_mode    = "workspace-write"

[profiles.readonly_quiet]
approval_policy = "never"
sandbox_mode    = "read-only"

Sandbox mechanics by platform

The mechanism LLMX uses to enforce the sandbox policy depends on your OS:

• macOS 12+ uses Apple Seatbelt. LLMX invokes sandbox-exec with a profile that corresponds to the selected --sandbox mode, constraining filesystem and network access at the OS level.
• Linux combines Landlock and seccomp APIs to approximate the same guarantees. Kernel support is required; older kernels may not expose the necessary features.
• Windows (experimental):
  • Launches commands inside a restricted token derived from an AppContainer profile.
  • Grants only specifically requested filesystem capabilities by attaching capability SIDs to that profile.
  • Disables outbound network access by overriding proxy-related environment variables and inserting stub executables for common network tools.

Windows sandbox support remains highly experimental. It cannot prevent file writes, deletions, or creations in any directory where the Everyone SID already has write permissions (for example, world-writable folders).

In containerized Linux environments (for example Docker), sandboxing may not work when the host or container configuration does not expose Landlock/seccomp. In those cases, configure the container to provide the isolation you need and run LLMX with --sandbox danger-full-access (or the shorthand --dangerously-bypass-approvals-and-sandbox) inside that container.

Experimenting with the LLMX Sandbox

To test how commands behave under LLMX's sandbox, use the CLI helpers:

# macOS
llmx sandbox macos [--full-auto] [COMMAND]...

# Linux
llmx sandbox linux [--full-auto] [COMMAND]...

# Legacy aliases
llmx debug seatbelt [--full-auto] [COMMAND]...
llmx debug landlock [--full-auto] [COMMAND]...