Skip to content

Safety and workspace access

zarlcode runs with your user privileges. It is a local coding agent, not an isolated hosted service: Build mode can edit workspace files and run processes. Treat a model’s requests as suggestions to review, particularly in an unfamiliar repository.

Plan mode is the safe default for investigation. The agent can inspect the workspace and produce a structured proposal, but it cannot edit files or run shell commands. Review the plan in the timeline or with Ctrl+P, then use Shift+Tab to enter Build mode only when you want execution.

This is a workflow guardrail, not a substitute for reviewing a change. Use the visible timeline and working set to understand what Build mode did.

The coding tools operate relative to the workspace you launched from. File reads, edits, searches, and directory listing are routed through the agent tool system; each shell command is tracked so it can be inspected or stopped instead of disappearing into terminal scrollback.

Open Ctrl+W to inspect changed files, diffs, and tracked processes. The working set can roll a file or turn back to its recorded checkpoint. Use Ctrl+F to see the files, skills, agents, and hooks that are visible to the session.

Shell policy and process guardrails apply to Build-mode commands. On supported Linux systems, sandboxing can additionally confine shell execution with Landlock. It is a useful layer, not a guarantee that removes the need for a suitably scoped user account or review.

Web tools, MCP servers, and other external integrations can access systems beyond the workspace according to their own configuration and credentials. Confirm which MCP servers and tools are enabled before giving a task authority over sensitive systems.

Sub-agents are separate runs owned by the parent task. They make focused research or verification easier to inspect, but they share the same trust boundary. Overlapping workspace operations wait in fair arrival order rather than racing; disjoint paths can proceed concurrently.

For the implementation detail behind these controls, see The tool system, Guardrails, and Sandboxing.