Agent Sandbox Dry-Run Guard

Rehearse risky AI agent file operations in a clone before touching the real disk

Problem

A developer let a fully-permissioned coding agent run tests unattended and came back to a wiped Desktop folder, an emptied recycle bin, and broken app icons; another had the whole Desktop deleted by an AI writing a plugin, with nothing recoverable. In both cases the agent issued destructive delete/move operations with no isolation or confirmation step, leaving the user to salvage afterwards.

Solution

A lightweight local guard that intercepts high-risk commands from agents (rm -rf, bulk moves, writes into system directories), clones the target paths into a temp sandbox, dry-runs the operation, and shows exactly which files would be deleted or moved before asking for one-click approval. Anything unapproved is blocked and a rollback snapshot is captured automatically.

Users

Indie developers and programmers running Claude Code, Codex, opencode or similar coding agents on real projects, often with auto-approve or full permissions enabled

MVP

Intercept and parse common dangerous shell commands; dry-run path-affecting operations and list affected files; one-click rollback snapshot for user-specified directories; local CLI wrapper that plugs into existing agents with zero config; incident log and export

Revenue

One-time desktop/CLI license; Pro subscription for multi-directory snapshots and team policy sync

Why now

Coding agents are now used daily but their permission model is still all-or-nothing, real data-loss incidents are already showing up in communities, and big vendors are unlikely to build fine-grained local guards soon