The Private-AI starter guide
A plain-English primer · about a 5-minute read
What a private AI actually is
Think of the AI everyone else uses as a payphone in the town square: you walk out, hand your conversation to a stranger behind the glass, and trust them not to listen, not to write it down, and to still be there tomorrow. A private AI is your own phone, in your own home. The conversation happens inside your machine and stays there. It works in a blackout. And no one can hang it up or change the rules on you.
The thing that got installed is a model — a single large file that holds everything the AI learned, a bit like a dictionary that already knows how to write and reason. Once it's on your computer, the part that does the thinking sits there with it. When you ask it something, your machine works out the answer by itself — the way it adds up a spreadsheet or edits a photo. No message goes out to the internet, so there's nothing to intercept and no account watching.
Who it's for
Plainly: most people only ever need the free app. If you want a capable assistant for everyday writing, thinking a decision through, or asking questions about your own documents — and you'd like all of that to stay on your computer — the free Garnet app is the whole answer. You don't need anything paid, and we'd rather tell you that up front than sell you something you won't use.
The paid side — a dedicated build — is for a narrower group: someone who needs the AI always on and ready, a model far larger and sharper than a personal laptop can hold, or a private system wired into specific, sensitive work. If that's not you, the rest of this guide is everything you need.
What it's brilliant at (and what isn't)
Short version: it's brilliant at the everyday things, and the cloud giants still edge it on the rare hardest questions. We'd rather you know that honest trade before you start.
Brilliant at, right on your machine:
- Everyday writing — emails, messages, rewrites, plain-English explanations.
- Thinking out loud — planning, untangling a decision, a journal that answers back.
- Your own documents — drop in a letter, a lease, a report and ask plain questions, knowing the file never travels anywhere to be read.
- Reading and summarising — long emails, dense PDFs, a stack of notes turned into the few lines that matter.
Still better in the cloud: the very hardest reasoning, and anything that needs live, up-to-the-minute web data. The giant assistants run on enormous machines in data centres; for the rare hardest questions they can still pull ahead. For nearly everything people actually reach for day to day, the one on your own computer answers right away — and answers only for you.
How big a model fits on your computer
Here's the one technical thing worth understanding, because it explains nearly everything else. To answer you, your computer has to load the whole model into its memory (RAM) at once. Bigger models know more and reason better — but they're heavier, and they can only be as big as your memory has room for. That's the whole trade in a sentence: more memory, a bigger model, a sharper Garnet.
Roughly, and honestly — these are guide-rails, not promises:
- 8 GB — a genuinely useful everyday model. Great for writing, rewriting, and reading your documents. This is where most laptops land.
- 16 GB — a noticeably smarter model that holds a longer train of thought. The sweet spot for most people.
- 32 GB and up — room for the larger, sharper models that handle harder reasoning without breaking a sweat.
Apple-Silicon Macs (the M-series) share memory cleverly between the chip and the AI, so they punch above their number — a Mac often runs a model that a same-memory Windows PC would struggle with.
How to get the most out of it
A few habits make the difference between a so-so answer and a genuinely useful one:
- Be specific. "Make this kinder and shorter, and keep the part about the deadline" beats "fix this."
- Give it the raw material. Paste the actual email, the actual paragraph, the actual numbers. It works best with something real in front of it — and since nothing leaves your machine, you can hand it the real thing, not a sanitised version.
- Think out loud with it. You don't need a clean question. Say "I've got three things tangled up and I can't see the shape of it" and let it help you sort the mess.
- Ask it to show its reasoning when the answer matters — "walk me through how you got there" — so you can check it. Local models, like all AI, can be confidently wrong; on anything that counts, treat it as a sharp first draft, not the last word.
- A roomier, newer computer runs a sharper Garnet. More memory means a more capable model — it's the one thing that most changes how well it answers.
Prove the privacy to yourself
You don't have to take a promise on faith. Once Garnet is on your computer, pull your internet cable (or turn off Wi-Fi) and ask it something. It keeps right on answering. That's the whole argument in one move: if it still works with nothing connected, then nothing you type was ever going anywhere. No account to hijack, no database of your chats to leak — because neither exists. If you'd like to see how we hold ourselves to that, here's exactly what we can and can't see.
To be precise, because we'd rather you know exactly what "private" means here: anything Garnet keeps — your past conversations, your settings — is kept on your own computer, the same way a saved document is. It's yours to read or delete whenever you like. "Private" doesn't mean nothing is ever written down; it means nothing is written down anywhere but your machine, where only you can reach it.
Will it run on your computer
Probably yes. Most computers from the last few years run Garnet comfortably — 8 GB of memory or more, on Mac or Windows. Not sure how much memory you have, or what it all means? The section above is the plain-English version, and the 10-second checker on the homepage tells you in plain words where you land: smooth, fine, or maybe go with a setup.
On an older or smaller machine it still runs — just with a smaller model, so expect quick and useful rather than its sharpest. If that's you, the checker will say so plainly.
Ten prompts to start with
Copy any of these straight in and swap in your own text where it says [paste]:
- Turn this messy idea into three clear bullet points: [paste].
- Write a kind but firm reply declining this: [paste].
- Explain what this paragraph from my lease actually means: [paste].
- I'm overwhelmed this week — help me make a simple three-day plan.
- Read this bill and tell me, in plain English, what I'm being charged for.
- Be my devil's advocate on this decision: [describe it].
- Rewrite this so a smart twelve-year-old could follow it.
- Here are my notes from a meeting — pull out the things I actually have to do.
- Help me draft a hard message to someone I care about, gently.
- Ask me five questions that would help you give me a better answer.
When you want more than the app
The free app runs a genuinely capable assistant on your own computer, and for almost everyone that is the whole answer. Some people need more: a dedicated, always-on system; a model far larger than a personal machine can hold; a private AI built around their specific work, their files, their rules. That's a build, not an install — a serious piece of engineering. The engineer who would architect it is the person you talk to, not a salesperson.
We'd rather you know the scale up front: a serious dedicated system is real hardware and real engineering, and the larger ones run well into six figures — a top-end private model can pass $250,000 once you count the machine it lives on. That isn't a price we'll quote you on a webpage, because it depends entirely on what you actually need. Most people who reach out don't need anything near that — they want a guided setup, or a mid-sized model tuned to their work, and they leave the call with a clear plan, not an invoice.
A build tends to make sense when one of these is true:
- The information you'd feed an AI legally or ethically can't touch someone else's servers — patient records, client files, privileged work, trade secrets.
- You want a model meaningfully larger or sharper than a laptop can run, living on hardware that's always on and always yours.
- You want it shaped around your specific work — your documents, your terminology, your rules — not a general assistant.
- You want one engineer who owns the whole thing end to end, and a system you keep even if we disappear.
If none of those quite fit, the free app is almost certainly enough — and the call will tell you that plainly, for free. The same conversation scopes all of it, from a guided install up to a dedicated build, so you don't have to know which one you need before you book.
Book a free scoping call Free, no card. You only pay if you decide to build.
How to get Garnet
The app is free and coming soon. If you just want the free app: join the waitlist and we'll email you the day it's ready to download — one email, the day it lands, nothing else. No account, nothing to subscribe to, your words stay on your own computer.
If you already know you need more — you want it set up for you, or a dedicated build around your work — book a free scoping call instead. No card; you only pay if you decide to build.
Not sure which is you? Start with the waitlist. You can always book a call later, and the call itself will tell you honestly whether the free app already covers what you need.
More: what Garnet is · what we can and can't see · the engineer who builds it · back to top.