
Some Files Should Never Touch the Cloud. Use Local A.I. Instead.

Google A.I. Is the Search Result. Is Your Website In It?

A Device to Use AI on Files You Wouldn’t Share With Cloud AIs.
The idea has been in my head since 2011. A thumb drive I could plug into any computer. A library computer. A laptop in a hotel. A work PC while traveling. Encrypted files on the drive, so if I lose it the files stay closed. A private place to work on those files. Secure ways to share those files or do research on them. Nothing left behind on the host computer.
I wanted four jobs on that drive. Work on the file with AI that is not in the cloud. Browse privately. Message privately. Encrypt the file so a lost device is not a leak.
I could not build it then. I would have needed a large team of developers. I decided to try it with the tools we have now. Three hundred [of my] hours later, it runs.
What it actually does
Plug it in. Open a document. Ask it to find the trends. Ask questions about the file. The A.I. model runs on the device. If the work is the kind of thing cloud models refuse, I can switch to an uncensored local model. Different models have different values. The one with the lock is the one I use when the conversation has to stay private.
Private browsing is on it. Tor if I want that path. Messaging is on it. Session is the default I recommend. Some people prefer SimpleX. I am not trying to hide. I want options to send patient files or sensitive financial data without wondering if that email was secure.
What it is not
It is not Claude Desktop with a nicer case. Claude Desktop still sends the work to the cloud. This one does not. The AI is on the USB. It has no cloud access. It works on the thing I tell it to work on. Nothing escapes it on purpose.
It is also not magic. If the host computer has spyware, that spyware can see the screen. A untrusted PDF can carry a prompt injection. After I read those risks I designed the tool so even a prompt injection does not get the browser and the files in the same reach.
Why a drive, not another app
I am not always on my own machine. I go to a customer’s office. The files live on a specific computer. I need to plug in, do the work, leave the result, and walk out. I do not want fifteen USB sticks for encryption, messaging, and a model. I want one environment.
There are other versions on the bench. A dual drive that can enable these features to a phone or an iPad. A larger portable SSD for heavy local jobs, like a folder of 100,000 scanned PDFs with handwriting that should never hit a cloud model. Those are personal builds for work I already have. Not an intentional product launch.
One of my users said I should make it available to others, so I am, by request only. Reach out.
The member takeaway is the same whether you ever touch this drive or not. If the file cannot go to the cloud, stop pasting it into a cloud chat. Describe the job. Build a local tool. Or use a local model on a machine you control.
- If you would not email the file to a stranger, do not paste it into a cloud model.
- Ask Claude to build a program that runs on the folder, not a prompt that needs the file uploaded.
- When you need a model on the file itself, run a local model on a machine you control.
- A plug-in drive is one way to take that environment to a client computer and leave nothing behind.
- Desktop AI apps still call the cloud. Do not treat them as local.
This post came out of a real conversation at the AI Essentials Roundtable, the small group I run for business owners who want to actually use AI instead of just reading about it. We meet every other week, screen-share real builds, and steal each other’s wins. If that sounds useful, details are here: tablelandpartners.com/ai_essentials_roundtable


