How it works

Everything it does, and where it does it.

PHNTM One is a full assistant, not a chat box. Here is the whole surface — what runs on the device, what stays there, and the one path that ever leaves.

Order PHNTM-One — Made to Order Verify the claims

The pillars

Eight things it actually does today

Every capability below ships now, on the hardware, at version 0.6.9. Nothing here is a roadmap promise.

01 · local intelligence

A real model, on the device

A Gemma 3 4B model answers from the box itself — with the internet unplugged. A smaller local model stands by as an automatic fallback, so the assistant keeps working even while the main model reloads.

gemma3:4b · num_ctx 8192 · local llama3.2:3b fallback

02 · memory

It remembers what matters

It builds a personal, on-device memory of your names, projects and preferences — and recalls them by meaning, not keyword. Correct it and the old fact is superseded; ask it to forget and it does. Everything lives in one place: the box on your desk.

SQLite + local embeddings · corrections · forgetting

03 · document intelligence

Give it your documents

Hand it your files and it indexes them locally for retrieval-augmented answers grounded in your own material — and it tells you when it is answering from a document versus from the model. Delete a document and it leaves the index too.

local RAG · source-aware · nomic-embed-text

04 · voice

Talk to it, out loud

Press TAB or tap the mic to speak; press again to stop. Speech-to-text and text-to-speech both run on the device. The microphone is live only while the dot is red, caps at 15 seconds, and there is no wake word — it is never listening on its own.

whisper.cpp STT · Piper TTS · tap-to-talk, no wake word

05 · phone + drop

Your machine, from your pocket

Scan the on-screen QR once and a private chat opens in your phone’s browser — no app store, no account, no cloud relay in between. PHNTM Drop lets you send text and files straight from the phone into the assistant, all across your own network.

QR pair · same-LAN · no relay

06 · follow-through

It closes the loop

Set one-off and recurring reminders, pause or skip them, and let it track open loops so nothing falls through. Reminders survive restarts — the schedule lives on the device, not in a browser tab.

one-shot + recurring · open loops

07 · trust

Built to be questioned

Watch the Wire lists every live outbound connection with the service that owns it — verify it yourself with tcpdump. The box watches its own disk, memory and services, heals stuck services, and reports what it did in plain language.

phntm privacy · live egress · self-healing caretaker

08 · ownership

You own the whole thing

It is a Debian Linux computer. SSH is documented and never removed. Take Everything exports your data, and owner-initiated updates roll themselves back if a self-test fails — nothing changes underneath you.

Debian · SSH · Take Everything export · self-test rollback

Two brains

Private by default. Boosted when you choose.

Private · default

All local, free forever

Nothing you say leaves the device. No account, no subscription, no cloud. The local model does honest work with real privacy — this is the reason the product exists, and it is where most people stay.

Boosted · optional

A frontier model, your key

Bring your own API key for dramatically more capable answers on a hard task. It is off by default and runs on your own credit. It is not private: your prompt leaves the box when it answers, and the device labels every answer that did. Local always stays the fallback.

Honest limits

Deliberate trade-offs, stated plainly

A 4B local model is genuinely useful, not a frontier model. It can be wrong — the device always shows which brain answered.
Local inference is not instant. A warm answer is typically 15–20 seconds; the very first answer after an idle spell takes about 90–106 seconds once, while the model reloads to give you the RAM back. Later turns are fast.
Voice is tap-to-talk by design. There is no wake word and the mic is never always-on.
Boosted is the single non-local path, off by default and clearly labelled every time it applies.
It runs within an 8 GB budget on a Raspberry Pi 5. That constraint is the point — if it runs here, it runs anywhere.
The product keeps evolving. Everything on this page is what ships today; the vision is labelled separately.

Your AI. On your desk.

Made to order, built by hand, burned in, and tested before it ships.

Order PHNTM-One — Made to Order

$549 · one-time · no subscription · ships within 7–10 business days after payment