Your AI. On your desk.
_Coming_PHNTM Umbra, the bigger brain · target Holiday 2026 →
Your AI. On your desk.
A private AI assistant you actually own. It remembers you, reminds you, reads your documents, and talks back. By default, nothing you say ever leaves the box.
Order PHNTM One · $549 12 reasons to own one ↓
Now accepting orders · ships within 7–10 business days after payment · free US shipping
100% local by defaultZero telemetryNo accountNo subscription, ever
Autonomy
It takes care of itself.
All four are in the current software (0.8, as of September 25). Self-repair and keeping score are early: they cover specific cases today, and more repair types are in development.
Available now
Self-heals
It watches its own disk, memory and services, restarts the ones that get stuck, and tells you in plain words what it did.
How it heals Available now
Autonomous Mode
Switch it on and Claude, using your own Anthropic key, acts for you from Telegram: it checks the device, sets reminders, drafts and sends email and manages your Google Calendar. It's off by default and not private: what you ask goes to Anthropic.
Available now · early
Repairs itself
Deeper faults are repaired as a transaction: snapshot, fix, verify, then keep it or roll it back. Today that covers a drifted configuration file, and the memory and conversation search indexes. Your own words are never rewritten from an index.
Available now · early
Keeps score of every fix
It records every issue and how each repair turned out. Where it has more than one allowed fix, it can try one out of the shipped order first, once each has been tried at least five times across at least three separate incidents and one has a clearly better record. Every unit starts with a blank record.
See how it looks after itself →
This is the builder's own unit, with his Claude key and Telegram connected, which is why the screen shows an amber “1 external connection” badge. In Private mode, the default, nothing you say leaves the box, and the device labels anything that does. How privacy works →
Gemma 3 4B, on the deviceRaspberry Pi 5 · 8 GB10.1″ touchscreenWorks with the internet unpluggedZero telemetryNo account$549 onceSSH includedHand-built · 24 h burn-inFree US shipping30-day returns Gemma 3 4B, on the deviceRaspberry Pi 5 · 8 GB10.1″ touchscreenWorks with the internet unpluggedZero telemetryNo account$549 onceSSH includedHand-built · 24 h burn-inFree US shipping30-day returns
12
12 reasons to own one
Every other AI sends your words to someone else's computer. This one doesn't.
PHNTM One is a small computer with a real AI assistant built in. It sits on your desk, learns your world, and answers from right there. Here's why people want one.
01
Your words never leave your house.
The model runs on the device. In Private mode, which is the default, your questions, memories and documents stay on the box. They aren't sent to us or to anyone else.
the box · private mode outbound: 0 bytes of what you say
02
Pay once. Own it forever.
$549once$20 a month, forever
No subscription and no account. Nothing to renew, and no company can switch it off.
Do the math 03
It remembers you.
It learns your names, projects, decisions and preferences, and recalls them by meaning. Correct it and it updates. Ask it to forget and it forgets.
It follows through.
One-off and recurring reminders, plus open loops it tracks for you. It nudges you on its screen and on your phone, and its schedule survives restarts.
Ask your own documents.
Send it a file from your phone, then ask questions about it. It answers from your document, says when it's doing that, and keeps the file on the box.
Works with the internet unplugged.
Pull the network cable and it keeps answering, because everything it needs is already inside.
Talk to it. It can't eavesdrop.
Tap to talk and it speaks back. There's no wake word, and the mic is live only while the red dot shows. Speech is processed on the device.
listening · tap to stop15 s max
whisper.cpp → on-device · no wake word · no cloud voice
08
In your pocket, no app.
Scan the QR code on its screen and a private chat opens in your phone's browser. No app store, no account, no cloud relay.
Phone + Drop 09
Don't trust us. Watch the wire.
Watch the Wire counts what your conversation sends (zero in Private mode unless you've connected Telegram or email) next to every byte the whole machine sends to the internet. You can check it yourself with tcpdump. We'd rather you check than take our word for it.
full conversation · from /proof
$ sudo tcpdump -i any -nn 'not (net 192.168.0.0/16 or …)'
IP x.x.x.x.123 > pool.ntp.org.123 · time sync
IP 0.0.0.0.68 > 255.255.255.255.67 · DHCP renewal
IP x.x.x.x.5353 > 224.0.0.251.5353 · mDNS
…and then silence.
10
It looks after itself.
It watches its own disk, memory and services, and heals the ones that get stuck. Nothing updates on its own: you apply an update yourself, and it rolls back if a self-test fails.
Your data, portable.
Take Everything exports all of it whenever you like. It runs Debian Linux, and SSH is documented and never removed. There's no lock-in.
If we vanish, it keeps working.
There are no PHNTM servers for it to depend on, so there's nothing to shut down. Your unit doesn't need us to keep working.
PHNTM servers it depends onnone
Account to sign in tonone
Remote kill switchnone
Keeps working if we disappearyes
$549once, no subscription
24 hburn-in on every unit, minimum
FreeUS shipping, tracked
30\-dayreturns · 1-year warranty
Do the math
Stop renting your AI.
A subscription never stops charging. PHNTM One is paid for once. Set what you pay for AI each month and watch three years play out.
Over three years you keep $171 PHNTM One costs less from month 28 on.
Your monthly AI subscription
$20/ month
$10$20 · typical plan$30$50
AI subscription, $20/month PHNTM One, $549 once
Total spent over 36 months: an AI subscription versus PHNTM One. Hover, or focus and use the arrow keys, to read any month.
See the numbers
| Month | Subscription, total paid | PHNTM One, total paid | | ----- | ------------------------ | --------------------- |
This compares cost, not capability. A 4B local model is capable, not a frontier model, and the big cloud models are smarter. You buy PHNTM One for privacy and ownership; the savings are a bonus. Electricity isn't counted. Boosted mode is optional; it uses your own Anthropic key, billed to you per use. Full comparison
Before you choose
How it compares.
The short version of each head-to-head. The full comparison has every row, including the ones we lose.
vs chatgptPHNTM One vs ChatGPTChatGPT is smarter and faster, and it runs on OpenAI's servers. PHNTM One is one payment, no account, and in Private mode nothing you say leaves the house. vs alexaPHNTM One vs AlexaAlexa is faster, cheaper and controls your smart home. In Private mode, PHNTM One keeps your words in your house and never listens for a wake word.
See it up close
A small machine with a big presence.
A 10.1″ touchscreen that greets you by name, shows what it remembers and what's coming up, and tells you plainly what it's doing. Tap any photo to see it full size.
AI-retouched photo 
AI-retouched photo 
AI-retouched photo 
AI-retouched photo
Real hardware · 21 s video
Photos are of the real device and its real interface, retouched with AI for lighting and background. The video is unedited.
A day with PHNTM One
A face on your desk. A chat in your pocket.
Set it up on the touchscreen, with no computer required. After that, talk to it at your desk or pick it up from your phone. Both are served by the device itself.
7:42AM · Morning
It keeps track of your life.
Tell it once and it follows through: one-off reminders, recurring ones, and the open loops you'd otherwise lose.
- Pause, skip or reschedule by just saying so
- Nudges on the screen and on your phone
- The schedule lives on the device, not in a browser tab
- Recalls by meaning, not keyword
- Corrections replace the old fact
- "Forget that" actually forgets
- No app, no account, no relay
- Answers say when they come from your document
- Delete the file and it leaves the index too
- Nothing you say ever leaves the device.
- No accounts, no subscriptions, no cloud.
- A capable on-device model that does honest work with real privacy.
- Works with the internet unplugged.
- Dramatically more capable answers when you want them.
- Runs on your own API credit, so you pay for what you use.
- Not private: your prompt leaves the box when it answers.
- Off by default. Local is always the fallback.
- The device, 10.1″ touchscreen
- Push-to-talk mic
- Mini wireless keyboard
- Power supply
- 256 GB microSD, model installed
- Quick-start card
phntm-one
all systems healthy private, on-device
illustration · sped up; real answers take tens of seconds
2:15PM · Afternoon
It remembers what you decided.
It builds a private memory of the people, projects and decisions in your life, and it takes corrections well.
phntm-one
memory · 100% local private, on-device
illustration · sped up; real answers take tens of seconds
8:30PM · Evening, on the couch
Drop in a document from your phone.
Pair once with a QR scan. Send a PDF from your phone and ask about it from the couch. It never goes through anyone's cloud.
illustration · sped up
The choice is yours
Three modes. You decide.
Most people never leave Private mode; it's the reason the product exists. If you want more power for a hard task, you can bring your own key. Flip the switch to see exactly where your words go. A third mode, Autonomous, is Boosted plus actions from your own Telegram, and it is off by default too.
Private default Boosted optional
PHNTM One · the box You Gemma 3 4B on-device model memory · documents · voice · all local Frontier model your API key leaves the box answer labelled PHNTM One · the box You Gemma 3 4B on-device model memory · documents · voice Frontier model your API key leaves the box answer labelled
Private: your question goes to the model inside the box and the answer comes back. Nothing crosses the line. Boosted: your question leaves the box for Anthropic, on your own key. The answer comes back labelled, so you always know.
Private · default
All local
Free, forever
Boosted · optional
Frontier model, your key
Your own API key
Who it's for
Made for people who'd rather own it.
Four kinds of people keep asking for one. Here's how each of them uses it, and where it falls short.
writersUnpublished work stays homeAsk your own drafts and notes; answers cite the file and page. familiesNo wake word in the houseReminders, shared memory and questions, push-to-talk only. small officesDocuments stay on the premisesOne price, no seats, no account, no cloud by default. tinkerersDebian, SSH, readable sourceReady out of the box, yours to open up. The BOM is public.
Why I built it
I wanted something that was mine.
Real photo · the build deskAn assistant that sits on my desk, knows how I work, remembers what I told it last week, and helps me get through the day — without sending a word of it to someone else's computer. Not a tab in a browser. Not a subscription that forgets me when I stop paying. A small machine with a face, that's there when I sit down and stays out of my way when I don't.
That's PHNTM One. It's a companion, not a search box: it learns your names, your projects, your reminders, your documents — and keeps all of it in the one place it belongs, the box on your desk. You own the hardware, you can read every line it runs, and you can unplug it from the internet and it doesn't care.
— Jacob, who built the one on his own desk first.
It's for you if
you want a private assistant you own outright, you like the idea of a device that gets to know you, and you'd rather check a claim than trust one.
It's not for you if
you need the smartest model money can buy every single time (a 4B local model is capable, not magic), or you want something that lives in the cloud and follows you across devices. No shame in that — but I'd rather you know before you buy.
Good to know
The local model writes about two words a second: a short answer usually takes tens of seconds, a long one can run past a minute, and the first word usually arrives within a few seconds. The first answer after an idle spell is slower, while the model reloads.
Arrived with questions?
The short version.
phntm-one · the short version
what it is
A small Linux computer with a private AI assistant built in: a physical appliance on your desk, not an app
price
$549 · one-time · no subscription, ever
local model
Gemma 3 4B (Q4\_K\_M), on-device · runs with the internet unplugged
hardware
Raspberry Pi 5 · 8 GB RAM · 10.1″ touchscreen · enclosed case, fan, speaker · Debian Linux
privacy
By default nothing you say leaves the box · zero telemetry · no account
ssh
Yes: SSH is documented and never removed
frontier model
Optional and off by default: bring your own API key, and the device labels every answer that left the box
in the box
The device (Pi 5 in a 10.1″ touchscreen) · USB push-to-talk mic · mini wireless keyboard · power supply · 256 GB microSD pre-loaded · quick-start card
ordering
Now accepting orders. Made to order in a limited production batch; ships within 7–10 business days after payment · order →
proof
Measured numbers and the commands to verify every claim →
Two choices
One is here. Umbra is coming.
Same promise, two shapes: the all-in-one you can order today, or a standalone mini PC being built for a much bigger brain.
available now · $549PHNTM OneThe all-in-one: Raspberry Pi 5, 8 GB, a 10.1″ touchscreen built in, and Gemma 3 4B on the device. Ships within 7–10 business days after payment. coming · target Holiday 2026 · $899 targetPHNTM UmbraA standalone mini PC with an 8-core AMD Ryzen and 24 GB, aimed at local models three to five times larger. You'll bring your own monitor and keyboard; the 5″ flip-up screen will be the face. Free waitlist, no deposit.
The vision
PHNTM exists to give people ownership of their intelligence.
We see a future where advanced AI doesn't have to live inside a distant data center, depend on a monthly subscription, or ask you to surrender control of your personal information. Intelligent systems should be able to run locally, learn alongside their owners, protect their privacy, and become more useful over time.
PHNTM is being built as a persistent, autonomous intelligence platform — combining local AI, memory, RAG, perception, automation, self-maintenance, and eventually robotics into one evolving system.
The goal is not to build another chatbot.
The goal is AI that lives on your hardware, understands your world, remembers what matters, takes meaningful action, improves through experience, and stays under your control.
Your hardware. Your data. Your intelligence.
For the skeptics
Check every claim before you buy.
We publish what other companies won't, because you should be able to verify the claims yourself.
proofVerify it yourselfMeasured numbers and the standard commands to check every claim on your own unit. bomEvery part, every dollarThe full bill of materials: $409.95 in parts and about $100 of real margin. transcriptsIncluding the wrong answersUnedited model output with annotations, plus the device's actual voice. compareWhat leaves your houseCompared with cloud speakers and chatbot apps, including the rows we lose. continuityThe dead-company testIf we vanish tomorrow, your unit won't notice, because there are no servers. unitsThe unit ledgerEvery serial ever built, with its own burn-in numbers.
Made to order · production batch 1
Own your AI.
$549once no subscription, ever
Batch 1
Every unit is built after you order, burned in for at least 24 hours, and tested. It ships within 7–10 business days after payment.
Order PHNTM One · $549 → Read the FAQ
Free US shipping 30-day returns 1-year warranty Secure checkout by Stripe
Not ready yet? Leave your email for build-log updates and news of the next production batch. No spam, and we never sell the list.
Keep me posted
You're on the list. We'll be in touch.
Or just email me directly with questions, orders, anything.