The problem: tools like Claude Code ask you to approve every file edit, every shell command. After hundreds of prompts you stop reading and just hit "yes", or even worse you reach for `--dangerously-skip-permissions` and hope for the best. Neither option is great.

yoloAI runs your AI coding agents (Claude Code, Codex, Gemini, Aider, OpenCode) inside disposable containers so that they can operate without permission prompts. Your files are never touched because the agent works on an isolated copy. When it's done, `yoloai diff` shows exactly what changed, and `yoloai apply` patches your real project.

Workflow:

    yoloai new fix-bug ./my-project --prompt "fix the failing tests"
    # agent works freely inside the sandbox
    yoloai diff fix-bug    # see what changed
    yoloai apply fix-bug   # patch your real project (either workdir changes or new git commits)
    yoloai destroy fix-bug # toss the container

This isn't just some Docker wrapper script. yoloAI handles the full lifecycle: copying your project into the sandbox, setting up git baselines, injecting credentials, managing agent state across sessions, and giving you a clean diff/apply workflow to review changes before they land. Your originals are protected until you explicitly choose what to keep.

Single Go binary. Works on Linux (Docker), macOS (Docker Desktop, Tart VMs, or native Seatbelt sandboxing). Profiles with user-supplied Dockerfiles for reproducible environments. Network isolation with domain allowlists. Five agents supported out of the box.

Early access: The core workflow is solid, rough edges remain. MIT licensed.

https://github.com/kstenerud/yoloai

I built this because I was tired of babysitting Claude Code through permission prompts. Now I can't live without it. Happy to answer questions about the design, the tradeoffs, or the sandboxing approach.