As local AI agents increasingly run shell commands directly on developer machines, a single bad rm -rf can wipe out hours of work.

I built SafeShell, a small tool that makes destructive shell operations reversible.

Before a command runs, SafeShell creates a lightweight filesystem checkpoint. If the agent deletes or modifies the wrong files, you can roll back instantly.

Example:

rm -rf ./build safeshell rollback --last

Key details:

No sandbox, VM, or container

No root access required

Uses hard links (fast, no data copying)

Compressed snapshot history

Single Go binary (macOS & Linux)

MCP support for agent-driven checkpoints

It’s meant to be a native safety layer for local AI agents, not a restriction mechanism.

Repo: https://github.com/qhkm/safeshell

Curious how others are handling filesystem safety for local agents, especially outside of containers.