shellfirm

Why we included this project

Most damage from an AI coding agent happens the same way it does for a person: one careless command, executed instantly. shellfirm sits between your terminal and the shell, checks each command against a library of over a hundred risky patterns covering git, filesystem operations, Kubernetes, Terraform, Docker, and the major cloud CLIs, and blocks anything dangerous behind a quick arithmetic challenge. The risk scoring escalates with context, so the test gets harder when you are connected over SSH, running as root, on a protected branch, or in a production cluster, and it suggests safer alternatives rather than just refusing. Teams can add their own safety rules to a project config file (additive only, so no one can weaken existing policies), and every blocked or allowed command lands in a JSON audit log. Because the same engine runs as a shell hook and as an MCP server, the guardrails apply whether the command comes from a human or from Claude Code, Cursor, or another agent.