When I ask people why they’d run AI on their own computer instead of a website, they say privacy. It’s the answer I hear most, and I think it’s the least interesting one.

Privacy is a defensive reason. It tells you what won’t happen to your data and nothing about what the AI can do for you. Look hard at what an agent does and the case for running it locally becomes a case about reach, meaning what it can see and work on. By that measure a local agent is more useful, by a lot.

The consultant who can only read what you fax

Take the sharpest analyst you’ve ever worked with and put them in a sealed room with no windows. They can’t see your desk, your files, your inbox, or the folder of receipts sitting on your laptop. The only things they know are the things you photocopy and slide under the door.

That’s roughly where a cloud AI agent sits. It runs on its company’s computers and reaches your life through what you carry to it. That means the text you paste, the PDF you upload, the accounts you connect, and sometimes context you already explained last week.

The intelligence in that room might be world class, and it still doesn’t help much. For most real work, a genius who can see only the three documents you happened to upload is less useful than a merely competent helper with every file the job needs in front of it.

How useful an agent is depends on what it can reach

The privacy framing hides how usefulness works. I think of it as three things multiplied together: how smart the model is, how much of your context it can see, and how much it may do for you.

A cloud tool maxes out the first and makes you work for the other two. You can feed it context one paste or one upload at a time, and each handoff is a little friction. Enough friction and you stop handing things over.

Multiply it out. A brilliant model that can see almost nothing of your world, and do almost nothing in it, makes a worse agent than a decent model with the files in front of it. Access multiplies everything else the model has.

I think this is why so many impressive AI demos feel useless on a Tuesday. The demo shows a model’s intelligence in a vacuum. Tuesday is you needing it to work on your actual files and accounts, the mess that exists, and finding that the smart thing sees only what you carry over to it.

The tasks you never bother to do

The real cost of the wall is the tasks you never attempt at all.

Say you have a Downloads folder with forty receipts in it and a spreadsheet to reconcile them against. With a cloud tool that means uploading forty files or copying numbers out one by one, and you’d rather not. So it doesn’t get done, or you grind through it by hand at 11pm. The setup costs more than the task, and the task dies.

On your own computer, the setup is one step you take once. Add the receipts and the spreadsheet to your agent’s knowledge and it reads them on your computer, checks one against the other, flags the three that don’t match, and writes you a one-line summary. The files stay in its folder, so next month’s receipts go in the same place and the job is one message.

When getting the files to the AI takes longer than the job itself, I skip the job.

Once that friction is low, the kind of work I hand off changes. It stops being the big one-time projects and becomes the small, repetitive, boring ones that were never worth the ceremony of uploading, pasting and re-explaining. A lot of the busywork in my life is like that, and it only becomes worth handing off when handing it off costs almost nothing.

The model can stay in the cloud while the agent runs on your computer

The obvious objection is that the best models run in the cloud. That’s true, and it doesn’t conflict with any of this.

You can keep the intelligence in the cloud and move the agent home, and that is how Archie works. The model runs on the AI account you connect, from any of seven AI companies, so the thinking still happens on that company’s own models. The free credits you start with run on our account, through our server.

The agent itself lives on your computer: the part that holds your context, reads the files you give it, remembers what happened last time and drafts what you asked for.

Your keys sit in your computer’s own password store, where we have no way to read them. Your files and its memory are kept in its folder on your computer, and what it reads goes to your AI company each time it thinks.

So you get both halves that the usual debate treats as a tradeoff. The reasoning is as good as the model you pick, and the agent works from your own computer with the files you give it and the accounts you connect. You pick its jobs with add-ons: skills for the work you ask for, and routines for the work that runs on a schedule.

That’s the part the privacy talk buries. Keeping your files on your computer means no server of ours holds them, though what your agent reads still goes to the AI company you picked. The better reason to want them there is that your agent works right next to them, and you stop sliding pages under the door.

Privacy is a fine reason to run an agent at home. For me, the reason it’s worth having at all is what it can work on from there.