AI Agents

Using AI coding angents like Claude Caude, OpenCode or Codex is encouraged at Renuo.

But when using it you must make sure to do so in a safe way!

LLMs are probabilistic - 1% chance of disaster makes it a matter of when, not if.

https://agent-safehouse.dev/

We recommend two tools for this:

nono

MacOS and Linux.

Install

brew install nono

Use

nono run --profile claude --allow-cwd -- claude

agent-safehouse

MacOS only.

Install

brew install eugene1g/safehouse/agent-safehouse

Use

safehouse claude

zsh

This is an example configuration @schmijos uses.

The assumption here is that there is a trade-off between security and comfort. Because the normal approval-based approach doesn't give any more real security we go explicitly for the unsafe but comfortable setting in the agents. For real security we have safehouse.

Only Claude offers a better trade-off with it's security reviewing agentic-hook called auto-mode since May 2026.

safe() {
    case "$PWD/" in
      "$HOME/p"/*) ;;
      *)
        printf 'Not under %s (PWD=%s). Continue? [y/N] ' "$HOME/p" "$PWD"
        read -r reply
        [[ "$reply" == [yY]* ]] || return 1
        ;;
    esac

  safehouse \
    --workdir="$PWD" \
    --add-dirs-ro="$HOME/p" \
    --add-dirs-ro="$HOME/.gitconfig" \
    --add-dirs-ro="$HOME/.gitignore" \
    -- "$@";
}
claude()   { safe claude --permission-mode auto "$@"; }
codex()    { safe codex --dangerously-bypass-approvals-and-sandbox "$@"; }
amp()      { safe amp --dangerously-allow-all "$@"; }
gemini()   { NO_BROWSER=true safe gemini --yolo "$@"; }