Introduction

I’ve been using AI coding assistants for a while now. Anthropic’s Claude Code is my current favorite, and I use it daily for tasks ranging from simple coding jobs to Windows OS automations, massive refactors, and full feature development. Sometimes I just want to let the agent loose—and if it’s on the right track, review its work when it’s done.

You can do that with the virtual environments that Anthropic or OpenAI provide (like Codex), but those are often restrictive. You’re limited by what these companies give you—features of the environment (can I run containers?), processing speed, and power—and you can’t see results as instantly as when these tools run locally.

In this post, I’ll show you how to leverage what’s known as Claude Code YOLO mode: a way to run Claude Code without permission prompts. It can do whatever it wants—run wild while it figures out the problem you’ve given it. I’m a firm believer that to unlock the full potential of these tools, you need a constrained sandbox where you can give them full permissions safely and see what they’re capable of.

Our objective will be to minimize permission interruptions and maximize autonomy—with as little guidance as possible.

Requirements

To follow this guide, you’ll need a Windows machine with Virtualization Enabled. You’ll also need Docker Desktop installed if you want to leverage Docker inside the virtual environment.

Enter WSL2

A virtual environment is the first thing that comes to mind. You could go full virtual machine with VirtualBox, but we can go lighter with WSL.

The Windows Subsystem for Linux (WSL) lets you run Linux distributions (like Ubuntu) inside Windows without needing a traditional VM. You can learn how to install WSL and pick a distro here.

For my setup, I picked Ubuntu 24.04.

# Install the distro
> wsl.exe --install "Ubuntu-24.04"

# Check it
> wsl.exe --list
Ubuntu-24.04

# Enter it
> wsl -d Ubuntu-24.04
kostas@ANDROMEDA:/$

Disable Automount

By default, WSL automatically mounts your Windows filesystem. Remember when I said we’d let this little beast loose? Giving it access to all your files—including Windows and Program Files—isn’t the best idea.

We’ll be very selective about what files it can see by default. We’ll achieve that by modifying two files: /etc/fstab and /etc/wsl.conf.

Disable Automount

Open /etc/wsl.conf (create it if it doesn’t exist) and set the following values:

[automount]
enabled=false
mountFsTab=true

[interop]
enabled=true
appendWindowsPath=false

This disables automount, ensures only the directories listed in fstab are mounted, and prevents WSL from appending your Windows PATH (which can have odd side effects).

Pick the folders you want to share

Now open your fstab file:

sudo nano /etc/fstab

The first two lines are required for the Visual Studio Code WSL extension so that you can connect from VS Code. If you don’t plan to use VS Code, you can omit them.

The third line mounts your c:/dev folder, which is where I keep my code. Adjust paths as needed.

c:/Users/konst/.vscode/extensions   /mnt/c/Users/konst/.vscode/extensions   drvfs   defaults,ro   0   0
c:/Users/konst/vscode-remote-wsl    /mnt/c/Users/konst/vscode-remote-wsl    drvfs   defaults,ro   0   0
c:/dev    /mnt/dev    drvfs    defaults,metadata,uid=1000,gid=1000,umask=022    0    0

Once you’re done, restart WSL:

wsl.exe --shutdown
# Wait a few seconds...

# Jump back in
> wsl -d Ubuntu-24.04

Providing Docker

My setup uses Aspire and TestContainers to run tests and projects, so Docker needs to be available.

First, install Docker Desktop.

Configure Docker to use the WSL2 engine

In Docker Desktop settings, enable these options:

Now restart your WSL distro as before, and the docker command will be available in your terminal.

Install Node.js in WSL

Next, install Node.js and npm so we can run Claude Code.

sudo apt update
sudo apt install -y ca-certificates curl gnupg
curl -fsSL https://deb.nodesource.com/setup_20.x | sudo -E bash -
sudo apt install -y nodejs

# Verify installation
node -v
npm -v

# You should see
v20.x.x
10.x.x

Install Claude Code in WSL

Now that we have npm, install Claude Code globally:

npm install -g @anthropic-ai/claude-code

Finally, Turn YOLO Mode ON

Moment of truth.

claude --dangerously-skip-permissions

After a few prompts, you’ll land in this menu with bypass permissions enabled:

Out of Scope

Git commits and pushes are out of scope for this post. To allow Claude to commit and push, you’d need to configure SSH keys and give it access to your repositories.

For now, you can configure .gitconfig so it can stage and commit locally, but sharing your SSH keys gives it more power than I’d recommend in YOLO mode. It’s safer to ask it to commit its work, then manually review and push yourself.

Conclusion

In this post, we learned how to unleash Claude Code’s full power locally by giving it unrestricted permissions inside a controlled environment.

Go now—fix a bug, refactor something, and watch it in action.

Handwritten by Kosta and slightly polished with ChatGPT 5.1.

As I finish the implementation of yet another feature with the help of Claude Code and without having to write a single line of code or tests, I am reminded once more of how fast our profession is changing. My codebase is a few thousand lines of code, mostly C# and Typescript. The feature was pagination applied to both backend endpoints and also frontend tables, so when you navigate to new pages, it requests them from the backend on the fly. I completed the task in under 2 hours and with all of the code, both backend and frontend, covered by tests. I have instructed Claude Code to use my coding style and that’s strict TDD 99% of the time. The side effect is a very well tested codebase.

AI coders love types and they also love feedback given by tests and failing builds(due to, you guessed it, strongly typed languages).

What you know today, may become out of date by the next month. Looking at the skill level of engineers out there there’s a very wide spectrum of AI literacy and AI-coding assistance expertise.

Starting from the developer who may refuse to use any AI tools at all, to the one who crawls Hacker News for the latest developments to this particular area. Do you remember when someone had Visual Studio in their CV as a skill worth making the interviewer aware of? I do. It was a big deal back then. It meant you didn’t rely on Notepad to write your code, it meant that you were efficient at your craft.

Got the analogy yet? Maybe let me try to drive it home.

A few years ago, Github Copilot came out. It was an interesting tool and it slowly captured a good portion of the market. To this day, there are developers who rely solely on Copilot. You can say that an evolution of this is ChatGPT/Gemini/Grok prompting. Perhaps, a successor to Stackoverflow. Here’s my problem, BLAH, give me the solution. Amazing. Once you get hooked to that, if you are curious enough, you might be wondering, OK, but how can I apply this to a codebase? How can I skip the back and forth to my browser and the inherent curse of being a clip board jockey?

Then, you may enter the agentic world and use Claude Code or a similar tool. But even then, there’s depth. Depth that is increasing by the day. From MCP tooling that you can hook up to it, to agentic personas you can create and leverage to do separate tasks … and so on.

The point is, if you are not keeping up with the latest developments in this area, you may find yourself left behind.

Getting Started

If you’re looking to dive deeper into agentic coding tools, I recommend starting with Claude Code: A Highly Agentic Coding Assistant from DeepLearning.AI. This course provides hands-on experience with the tool I mentioned and covers the fundamentals of working with AI coding assistants.

For advanced usage tips and techniques, check out this comprehensive guide that covers more sophisticated workflows and best practices.

So do try to keep up. It’s a brave new world, and it’s moving fast.