I love the feeling of using Claude Code with --dangerously-skip-permissions on, but it always made me feel a bit uneasy. It's not that hard to trick Claude into doing bad things, and with web search enabled especially, the odds of Claude running into prompt injection attacks or accidentally downloading and running a virus don't actually seem all that low.

Running Claude Code without web search and constantly having to respond to it's permissions prompts is so slow though! I wanted something that let me get the best of both worlds, so I (well, Claude mostly) ended up building cco, which runs Claude Code in a customized docker env that gives it access to all it's tools, but also keeps it safely away from the rest of your computer.

So far it's neatly solved my yolo mode misgivings, maybe it can solve yours too. It doesn't guarantee your computer will be 100% safe, but running Claude Code inside cco should definitely be safer.