Agent Sandbox Gate

Dry-run every dangerous agent command in CI before it runs

Problem

Coding agents delete home directories, overwrite configs, and fire off destructive commands. Today it's all-or-nothing: full trust or full block, with no "show me what this would do" layer — so long unattended runs feel like a gamble.

Solution

A configurable gate that routes agent shell/file operations through declarative policy checks. When a dangerous pattern matches (rm -rf, dotfile overwrite, writes to credential dirs), it pauses, shows a dry-run impact summary plus a suggested safer form, and lets you approve in one click.

Users

Developers letting agents run unattended, eng leads rolling agents out to teams, self-hosted agent platforms

MVP

Intercept and parse agent shell/file ops; built-in dangerous-pattern ruleset with custom YAML; readable dry-run impact summary for matched ops; one-time approve or permanent allowlist; local CLI + hooks, no backend required

Revenue

One-time local license ($29-49) + team policy-sync subscription; a rules marketplace as a follow-on

Why now

Agents are moving from autocomplete to autonomous execution, and home-dir deletion stories are turning into a top fear; rule engines and hooks play to indie strengths and don't depend on model capability