Guardrails that stop an agent before it pushes to main

Coding agents run shell commands. Build a shell shim that sits between an agent and the terminal, blocks the dangerous ones, and asks a human for the borderline ones.

Requirements

A wrapper the agent is pointed at instead of bash; every command passes through it

A rule file (YAML or JSON) of deny patterns, allow patterns and "ask" patterns

Deny by default for: pushes to protected branches, rm -rf outside the repo, anything touching ~/.ssh or .env

"Ask" opens a local prompt with the command and a 60-second timeout that defaults to deny

An append-only audit log of every command, its verdict and who decided

Bonus Points

A dry-run mode that shows what a transcript of agent commands would have been blocked

agentssecurityclishell

Free stage

MidFree

Starter repo

You connect the GoShip MCP server to Claude Code or Cursor. It records your prompts, tool calls and test runs, and that record is what gets scored.

Discussion (0)

No comments yet

Post a question about the brief, or a tip for whoever tries it next.

▶Sign in to comment