# sandshell

> Your AI coding agents ship with sandboxes. sandshell makes sure they're actually on.

**Status:** Released | **Language:** Bash | **License:** MIT
**Repo:** https://github.com/liwala/sandshell

## The Problem

Every AI coding agent ships with a sandbox primitive, but they all default to off, narrow inconsistently, and silently regress between sessions. You'd have to re-derive the same threat model for each one — by hand.

## What It Does

- Detect every AI agent and sandbox primitive installed on the host
- Audit findings by severity (critical, high, medium, low) across agents
- Apply safe-default configs to Claude Code, Codex CLI, and Gemini CLI
- Drift detection between sessions so quiet regressions surface
- Per-session Bash audit trail for Claude Code
- Interactive permission pruner for over-broad allowlists
- verify subcommand for CI / pre-commit (exits non-zero on findings)
- Project-scope configs you commit to git for your team

## Quick Start

```bash
git clone https://github.com/liwala/sandshell ~/sandshell
export PATH="$HOME/sandshell/bin:$PATH"

sandshell detect            # what agents do I have?
sandshell audit             # what's risky right now?
sandshell install-agent all # one-time agent guidance install
sandshell apply             # write safe defaults
sandshell drift             # later: what regressed?
```