In development · Reservations open

The things you shouldn't still have to think about.

Alfred is a private AI Chief of Staff: one secure memory across your whole life, a briefing every morning, and nothing consequential done without your say-so.

Runs on hardware you own Private persistent memory Approval before action

How Alfred Core is built

A finished Alfred Core: a compact, sealed aluminium machine with a single cable attached.
01 — The box

It starts as hardware you own

Alfred isn't an account on somebody else's servers. It's a machine that lives in your home or office, and everything private begins and ends inside it.

02 — The logic board

The model proposes. The policy engine decides.

Alfred is a permissioned orchestration system, not a chatbot with your passwords. Language models interpret and suggest; deterministic software controls what is actually allowed to happen.

03 — Local inference

The routine thinking happens here

Classification, extraction and transcription run on your own machine. Only the heavier reasoning is routed out — and only to the services you approve.

04 — Persistent memory

The part that makes it worth having

Identity, preferences, relationships, commitments, episodes and procedures — six kinds of memory, encrypted on your disk, building the picture no single app or assistant currently holds.

05 — Sealed and audited

Every consequential action is on the record

Least-privilege access to each connected service, and a tamper-evident ledger of everything Alfred does on your behalf — reviewable line by line, whenever you want.

Assembled

One machine. One owner.

Plug it in, and Alfred starts in observe-only mode — reading, learning, and earning its authority one stage at a time.

What it is

A private office, in software

As a life gets more complex — more property, more advisers, more travel, more people who need an answer — you become the integration layer holding it all together. Nothing you use remembers the whole picture.

Alfred is built to be that memory and that coordination layer: it reads what comes in, understands what it means for you, remembers it permanently, prepares what it can, and asks before it acts. Not a chatbot, not a calendar app, not a concierge service — one operating layer that absorbs pieces of all three.

  • Proactive — it tells you your passport expires in seven months before you wonder.
  • Quiet — it suppresses noise rather than generating more of it.
  • Explainable — every recommendation shows why it appeared and where the fact came from.
  • Evidence-based — it never reports "done" without proof it happened.

The full capability set

Today · Thursday
ScheduleLondon 14:00 — flights, driver and hotel confirmedPrepared
DecisionContractor quote for the Marbella roof: two optionsNeeds you
MoneyVendor invoice matched to the agreed scopeReady to approve
PeopleYou promised Patrick a follow-up on TuesdayDrafted
RiskProperty insurance renews in 21 daysFlagged
HandledCar service booked, warranty documents filedDone

Illustration of the Today screen. Interface in development.

Capability

Ten jobs, one memory

Every module shares the same lifecycle — capture, understand, remember, prepare, approve, execute, verify — so nothing is stranded in its own app.

The daily briefing

What's happening today, what needs a decision, what changed overnight, what Alfred already handled, and what's coming that you should know about.

Universal inbox

Every incoming signal sorted into one action — decide, approve, pay, reply, do, read, note, or ignore — instead of one undifferentiated pile.

Personal memory

Identity, preference, relationship, episodic, commitment and procedural memory, working together and attributed to their source.

Calendar intelligence

A Thursday meeting in London implies flights, a hotel, transfers, documents and a return time. Alfred connects those dots for you.

Relationships

Who you haven't spoken to in six months, what you last agreed with them, what you gave them at Christmas, what you owe them now.

Document vault

Passports, policies, contracts, warranties and invoices — read, classified, searchable, with expiry dates and obligations extracted.

Home & property

A profile per property: contractors, utilities, access, inventories, maintenance. "Prepare the house for Friday" becomes one instruction.

Travel

Your airlines, seats, hotels, dietary needs, drivers and documents, applied without being restated every trip.

Human escalation

When something needs a phone call, a negotiation or a person on the ground, it goes to one — through the same interface.

Every module in detail

Control

You decide how much authority it has

Five escalating modes. Nobody is ever moved up a mode they didn't choose, and Alfred never takes unrestricted authority over money, security or confidential matters — at any mode.

01

Observe

Reads and analyses only. No drafting, no action.

"Your insurance renews in three weeks."
02

Prepare

Prepares work but never executes it.

Drafts the reply. You send it.
03

Assist

Executes only after your explicit approval.

You tap approve; Alfred sends it.
04

Trusted routines

Auto-runs only the specific routines you pre-approved.

"Book my usual airport driver whenever this trip appears."
05

Controlled autonomy

Operates inside explicit limits, vendors, categories and budgets you set.

Approves known-vendor invoices under your threshold.

How privacy and control work

Principles

Built to be trusted, not to be addictive

6 Kinds of memory
5 Autonomy modes
100% Actions logged
0 Data sold, ever

The measure that matters is hours of attention returned to you each month. Alfred is one of the few products that should aim for you to spend less time inside it.

Where we are

An honest status report

Alfred is in development. The prototype is being built and dogfooded on a real life before anyone else is asked to trust it with theirs.

Now

Foundation

Local stack running on the first Alfred Core. One mailbox and one calendar connected read-only. The Today screen built against the memory layer.

Next

The core loop

Meeting capture and commitment extraction, drafted replies with sources attached, the approval centre, and a document vault that reads expiry dates.

Then

Demo-ready

A desktop app and an iOS companion, running a real life reliably for weeks at a time — the bar before a single private install is offered.

After that

First private installs

A small number of design partners, each set up in person, on a progressive-trust ramp from observe to prepare to approve.

The full roadmap

Dispatches

Build notes, roughly monthly

What got built, what broke, what we learned from the first private installs. No launch hype, no forwarding your address to anyone.

One email a month at most. Unsubscribe in one click.

Two ways to be early

Reserve a place in the first production run — no payment taken, no obligation — or request the investor brief if you build things for a living.