I gave Claude Code admin access (in a sandbox)

Superuser permissions in a self-contained Docker container

💡 Remember you can get free copies of my books here.

I love creating things. Writing code has always just been a means to do that. Claude Code (and tools like it) have allowed me to prototype software in a CRAZY was over the past months.

What I like to do:

1. Open the Claude App in voice mode and say: “I am about to ramble an app idea at you, your job is to write a structured specification markdown file that I will feed into Claude Code. <and then I ramble for 5 minutes>.

2. Create a folder on my laptop, plop in the specification file, open Claude Code and say “given what you know about my preferences, rewrite this specification file to suit my coding style”

3. Tell Claude “implement the specification”

Here is the problem:

Either I need to pay attention, or I enable the “dangerously skip permissions mode”

And since I have an actual job and a todo list the size of a contract with a crossroads demon, I don’t want to hand-hold Claude.

What’s a nerd to do?

Permissions are more like guidelines

And since Claude can run shell commands, there is no way in hell I will tell it to do whatever it wants. There are too many sensitive things on my laptop, from cryptographic keys that nobody should ever see, to files I don’t want to lose under any circumstance.

The “rules” you give Claude Code are often not as deterministic as you think. Much of the Claude settings files are more like pinky promises than hard security limits. That is why skipping permissions is explicitly labeled as “dangerous”.

It’s like what Captain Barbosa says about the Pirate Code: “more what you’d call ‘guidelines’ than actual rules”.

Infinite permissions in a limited container

I’m a big fan of Docker. It allows you to create pseudo virtual machines that create little isolated environments that you can give permissions as you please. So last weekend I thought, what if I take Claude and:

• Give it superuser access (passwordless sudo access)

• Enable “dangerously bypass permissions”

• Give it a system prompt that says “do everything without asking for confirmation”

• But put it in a docker container that only mounts the current directory as read/write

If I mount the host Claude folder I could basically wrap Claude in a safety blanket that means it can just go yolo and run with anything I tell it to do

Enter: Sir Claudius.

It ended up taking a few hours of experimentation, but I now almost exclusively use Sir Claudius, which I call as a command line utility called “claudius”.

My above workflow of creating a specification through audio and then letting Claude play now ends with running: “claudius yolo”. Which runs claudius in Yolo mode, meaning it:

• Enables --dangerously-skip-permissions

• Adds a system prompt saying “you should do as much as possible without user interaction”

• Tells claudius to commit but not push it’s work (it couldn’t even do that because your host ssh keys are not available in the Docker container)

I’m having crazy amounts of fun telling it to make little projects and just letting it run in vibe mode.

It’s completely useless for my actual job, but for low-effort side projects it is AMAZING.

Some examples

What did I make claudius do this week?

• claudius sandbox yolo (no file access, autonomous): give it my Porkbun API keys and tell it “find me 100 available one-word domain names, I am looking for a domain to host my pet projects, it should be easy to say and write”. It took 3 hours to run and I’m window shopping the domains now.

• claudius yolo: make me an ereader app that uses an OpenAI API key to translate any ebook I upload to a language of my choice, but let me select the language level. Example: take in a Shakespeare play but translate it to kid-level Albanian (which I’m trying to get better at). It took 30 minutes, it works and I’m looking at hosting it now

• claudius mudbox yolo (read only access to files): brainstorm names for this app, feel free to run the app and click through it with the Playwright MCP

It’s amazing what you can let AI do if you don’t need to worry about it breaking things.

How to try it

If you want to enlist Sir Claudius to your laptop, all instructions can be found at this Github repository I maintain. Installation takes 5 seconds and it uses your existing Claude auth (Claudius is just Claude with an extra cape).

💡 Remember you can get free copies of my books here.