Autonomy · PHNTM One

A computer that looks after itself.

PHNTM One checks itself every 90 seconds, repairs with an undo, finds the cause before it treats the symptom, and tells you every week what it did, in every mode, with the internet unplugged. It improves only when you approve. And when you want it to act for you, Autonomous mode waits for your yes before it sends an email or changes your calendar.

Works offline18 bounded repairs7 with an undoCode changes only with your OK
The self-maintenance loop Five steps in a ring: detect, decide, repair, verify, record. At the centre, the undo for configuration and index repairs: if any check fails, the device puts back exactly what was there and checks that too. undoconfig · index 01 Detectsomething died, driftedor filled up 02 Decidean allowed fix, inshipped order unlessits record clearly differs 03 Repairsnapshot, stage,apply 04 Verifycheck again,live 05 Recordkeep score ofhow it went
Self-maintenance

Four habits, each one bounded.

Local self-care paths do not require a cloud model. Activity records show available evidence; Home’s This week card counts automatic repairs, incidents noticed and fixes waiting for your approval.

About every 90 seconds

It heals.

A supervisor checks services, the model server, memory, disk, temperature, the network, the screen, its own configuration files and the indexes it builds from your memories and conversations. When a check confirms a fault, one of eighteen repairs may run.

Each repair is limited: no more than three restarts of one service in half an hour, no more than two rollbacks a day on any one thing, and no automatic retry for six hours after a rollback. Beyond the eighteen it only does housekeeping, like trimming logs when the disk runs low.

7 of the 18

It repairs, with an undo.

The seven repairs that touch files or indexes run as a transaction: snapshot, stage, check, apply, check again live. The change is kept only if every check passes.

If not, it puts back exactly what was there, checks that too, and records the outcome under Device → Activity. Your own words are never rewritten from an index; the index is rebuilt from them.

Cause, not symptom

It looks for the cause.

It can tell when a fault is the cause and when it is a consequence of something else, or say honestly that it cannot tell. A repair aimed at a symptom is withheld when the evidence says the cause is elsewhere.

When it can't answer a question about its own health, it can propose what evidence would settle it.

Learns on your unit

It keeps score, and rehearses.

Every incident and every repair outcome is recorded. Where it has a choice of fixes, it prefers what has worked on your unit. On quiet days, once it has proved its own safe window, it breaks something built to be broken and watches the real repair path handle it.

It only ever chooses among fixes it was already allowed to make, and it keeps score of how each repair has worked on this unit.

The eighteen

Every repair it may make, and its limit.

The same table is in the manual on the device. Nothing outside this list is ever repaired on its own.

RepairWhat it fixesUndo?Limit
Restart the assistantthe chat service stopped answeringno (a restart)3 per 30 min
Restart the model serverthe local model server stopped answering, or grew past its baselineno3 per 30 min
Restart the Telegram bridgethe remote control plane went downno3 per 30 min
Re-warm the modelsa model dropped out of memoryn/a (a reload)load capacity
Relieve memory pressurememory critically low for two checks in a rowno3 per 30 min
Restart the facethe screen stopped paintingno3 per 30 min
Re-run the boot self-testthe start-up check had not passedn/a (a check)3 per 30 min
Rejoin Wi-Fithe device is off a network it knowsno (changes no credential)3 per 30 min
Restore a configuration filea shipped session file drifted from its known copyyes: verified, undone if wrong3 per 6 h · 2 rollbacks a day
Regenerate the public-facts filethe non-secret facts file drifted from its generatoryes3 per 6 h · 2 rollbacks a day
Reconcile the memory indexthe search index disagrees with what it rememberedyes3 per 6 h · 2 rollbacks a day
Rebuild the memory indexthe search index is too far gone to reconcileyes1 per day · 2 rollbacks a day
Reconcile the conversation indexthe conversation index disagrees with your kept messagesyes3 per 6 h · 2 rollbacks a day
Rebuild the conversation indexthe conversation index is too far gone to reconcileyes1 per day · 2 rollbacks a day
Rebuild replayed historythe device's own derived history diverged from its recordyes3 per 6 h · 2 rollbacks a day
Restart the face overnightthe screen's browser has grown past its memory line (between 3 and 4 in the morning, while nobody is talking to it)noonce a night
Reload the model runnerthe chat model is mostly sitting in swap after other work squeezed it outn/a (a reload)1 per hour, only while idle
Restart the device (last resort)the local AI, the chat service or the background process has been down for 30 minutes and its repairs are used up; you are told first, with how long until the restart; never in the first 30 minutes after boot, during an update or other maintenance, or while an answer is being writtenno (a reboot; evidence is saved first)1 per day
Getting better

It improves. You approve.

When a better way to do something is handed to it, the change goes through the same steps every time, and the last step is always yours.

1

Built in a sealed room

A small change is built and tested in a workspace on the device that can't reach the running assistant, your data or the network.

2

Judged, not assumed

It is built and tested in that sealed workspace; nothing it does there changes the running device.

3

You say yes

Owner-approved signed updates are a documented path: the Update Center checks a signed update and installs it only with your password.

4

Watched, then kept

Install applies the update, runs the boot self-test and rolls back automatically if the self-test fails; a rollback that cannot verify itself is reported as needing recovery.

This week, on Home

Home’s This week card counts questions answered, memories saved, notes written, documents added, reminders delivered, automatic repairs, incidents noticed and fixes waiting for your approval over about the last seven days. Counts only; with no battery clock, “a week” follows the device’s own clock.

The stop rules

Every repair has a fixed budget: for example no more than three restarts of one service in half an hour, and no more than two rollbacks a day on any one thing, with no automatic retry for six hours after a rollback.

Autonomous mode

Message it. It acts, after you say yes.

The third mode, next to Private and Boosted. It is Boosted plus actions you ask for from your own Telegram. It is off until you turn it on, it uses Claude on your own Anthropic key (so it costs you money), and turning it on from a phone needs your device password.

What you can ask it to do
  • Check how the device is doing
  • Set, list and cancel reminders, and show your open loops
  • Search and read your email, draft a message or a reply, and send it
  • List your Google Calendar, and add, change or delete events

Only your own paired Telegram account, in a private chat, is listened to. On the screen and the phone app it never sends email or changes your calendar; those actions come only from Telegram.

How it asks first
  1. Every email is shown to you in full before it's sent (recipients, subject, body), and every calendar change is spelled out before it happens.
  2. Nothing is sent or changed until you reply yes. Anything else drops it, and it expires after ten minutes.
  3. The approval is bound to exactly what you saw. If the draft or the event changed since, it refuses.
  4. A forwarded message can never approve anything.
What leaves the box

Not private, and it says so at setup. On a Telegram turn, what you ask, the recent messages in that chat, your and the assistant's names, your language and time zone, your recent email and calendar listings, and the results of the actions it runs (which can include full email bodies) go to Anthropic. Your messages and its replies travel through Telegram. The email and calendar actions you ask for, searching, reading and listing included, reach your email provider and Google Calendar; only sends and changes wait for your yes.

Supported sends (calls to Anthropic, emails sent, calendar requests) are recorded in the device's own egress ledger, which is not a complete network history. Switch back to Private, and disconnect Telegram, email and calendar, and conversation content stops leaving; clock sync, DHCP and mDNS continue.

The boundaries

What self-maintenance never does.

Touch your accounts.Self-maintenance never reads or acts on your email or calendar. If Telegram is connected, it may send you a notice about what it did.
Browse the internet.Healing, repair and keeping score all work with the router unplugged.
Change itself without you.Updates and improvements are yours to approve: you tap Install on a signed update, and it attempts a verified rollback if the self-test fails.
Act on an AI model's decision.No model can start a repair or a restart. There is no tool for it.
Hide what it did.Fixes, undone repairs and budget stops show under Device → Activity.
Ready when you are

An AI that keeps itself running.

Order your PHNTM One. Each unit must complete a 72-hour burn-in before shipment. $799 once, with no required subscription for local use.

$799 one-time · ships within 6–8 weeks after payment (EU: pre-order) · free US shipping · ships worldwide · 30-day returns