Help

Everything works. Nothing here depends on Google Play Services, Firebase, or any Google library — there is no push notification service, no analytics, no Google font, and no Google-hosted script anywhere in the product. The app installs from Huawei Browser or any Chromium browser exactly as it does elsewhere.

That is not a special accommodation. The product simply never depended on those things, because it talks to you through a messenger you already have rather than pushing notifications at you.

If it can send a text message, yes. Set your channel to SMS and your agent talks to you the way anyone else would. You lose the app, the voice button and the share menu, and you keep every actual capability: it still reads your calendar, chases your invoices, remembers what you tell it, and answers you.

This works because the thinking happens on the computer you leave on, not in your hand. The phone is a keyboard and a screen.

Every page still works. The things stored in your browser are conveniences — which theme you picked, how far through the setup wizard you got, whether you dismissed the install prompt — and each one is written so that being refused loses the convenience and nothing else.

That is tested rather than assumed: the release checks render every page with storage made to throw instead of return, and with every optional browser feature removed.

The reading pages — this one, privacy, terms — are plain HTML and work completely. The demo and the dashboard need JavaScript, because they are applications rather than documents.

Nothing on the site is fetched from another company’s servers, so there is no third-party script to fail, no font service to be slow, and no CDN to go down. The only outside requests any page makes are exchange rates and your own spreadsheet, both on the money dashboard, and both only once you have set them up.

A computer you can leave switched on, and Node 18 or newer. That’s the whole list. It’s one file with no dependencies, so a laptop, an old desktop or a £50 mini-PC all work equally well.

You’ll also want a key from an AI provider. Anthropic, OpenAI, Google and Microsoft all give you one when you sign up. If you’d rather not use any of them, Ollama runs a model on your own hardware and costs nothing.

No. The Setup Wizard asks you questions and writes the config file for you. The only thing you type into a terminal is node agent.js, and the guide shows you exactly where.

If something goes wrong, node agent.js --doctor checks every connection in turn and tells you in plain words which one is broken.

About fifteen minutes if you already have an AI provider account, closer to thirty if you don’t. Choosing which messenger to use is the only decision that needs any thought — everything else can be changed later without starting again.

Thirteen are built in: WhatsApp, WhatsApp Business, Telegram, Matrix, Slack, Discord, Messenger, LINE, Viber, Microsoft Teams, Rocket.Chat, Mattermost and SMS. You can run more than one at once against the same agent and the same memory, so a message you send on Telegram is understood by the same assistant that answered you on WhatsApp an hour earlier.

Telegram and Matrix are the easiest to set up — both work by the agent calling out, so neither needs a public URL, and Telegram takes about two minutes with no business account or review. If you just want it working today, start there. Rocket.Chat and Mattermost matter if you like this product's ethos: they are messengers you can run on your own machine, next to the agent, so the whole conversation never leaves your house.

iPhone, iPad, Android, Mac, Windows, Linux and ChromeOS, all from the same link. There’s no app store involved — you open the page and choose "Add to Home Screen" or "Install".

On Android, Windows, Mac and ChromeOS it also joins the device’s own share menu, so you can send a receipt photo or a link straight to your agent from any other app. Apple doesn’t allow that on iPhone yet.

Yes, and their phone doesn’t need to be smart — if it can send a text, it can talk to the agent. Add them as one of the people on your plan (with their yes: their chats are their own, like every seat), and set what helps: “remind Mum to take her tablets every day at 9”, appointment reminders, a message that answers kindly when she texts back. You manage everything from your own phone; she never installs anything or remembers a password.

The honest limits, because this matters more here than anywhere: it is not an emergency service and not a medical device — no fall detection, no health advice, and a daily reminder is not the same as you ringing her. It makes the everyday easier; it does not replace you.

When the “Tell a friend” box appears on the front page, yes — and it’s the simplest one we could design: your friend types the code shown there at checkout (it takes a percentage off their first purchase), and answers one question — “Who told you about Sorted?” — with your email. For every friend who buys and names you, a month of your plan is credited to you. If the box isn’t on the front page, the scheme is switched off at the moment.

How it stays honest: the code is enforced by the payment provider — once per person, first purchase only — not by this website, which has no accounts and tracks nothing. There are no referral links to be tracked by, no points, and no leaderboard. One code, one question at checkout, one thank-you.

Yes, two ways. The demo on this site runs in your browser with no setup and no card. And when the trial box is showing on the front page, there’s a 24-hour full-access trial: one message activates it, you use the real agent in your own messenger for a day, and it expires by itself unless you choose a tier.

Yes, whenever you like. It answers to Robin to begin with. Say your name is Sam and that’s it — no restart, no config file. You can change it again as often as you want.

Three tones: warm, brisk and playful. Say tone brisk to switch.

Worth being clear about what this does and doesn’t change: it changes the wording only. A brisk agent and a warm one give you exactly the same facts, take exactly the same actions, and refuse exactly the same things. Tone is not personality-as-behaviour, and it never softens bad news.

From what you tell it and what you choose to connect. It keeps a memory file of your preferences, standing orders and open loops — sizes, budgets, delivery habits, focus hours — and corrects itself each time you correct it.

Ask it what have you learned at any point and it shows you exactly what it thinks it knows and how confident it is. Nothing about that is hidden from you.

Because it starts empty and refuses to guess. The part that learns your habits lives on your machine and won’t act on a pattern until it has seen around twenty-five examples, has at least five of the kind it’s considering, and is confident enough to be right. Until then it asks instead of assuming.

That’s a deliberate trade. An assistant that starts guessing on day one feels impressive for about three days and then does something you didn’t want with something you can’t undo. Waiting is the better failure.

Yes, but only when there’s genuinely something to say. It watches for clashing appointments, journeys too tight to make, things you need to leave by, and open loops that have gone quiet. If none of those apply it stays silent rather than manufacturing a daily update nobody reads.

It asks before anything it can’t undo, and keeps a plain list of everything it has done so you can look back. Deletions go to your bin rather than vanishing.

There are five things it will never do off its own initiative no matter how confident it gets: spend money, send email, delete files, change its own configuration, or escalate to someone else on your behalf. Those always wait for you.

Say settings and it lists them in the chat. There’s no separate control panel to learn. Permissions shows what it’s currently allowed to do, and you can change any of it in the same conversation.

Yes, and it is switched off until you say journal on. Off is the default on purpose: most people don’t want a journal, and a feature like this offered eagerly feels like being handled.

Then journal followed by whatever is in your head puts it in a file on your machine and nothing else happens. Half the value is that sentence — a worry in your head is shapeless, and a worry in a sentence has edges.

Say reflect and it reads your recent entries back and says what it notices, especially the thing you mentioned once and moved straight past. Decide followed by whatever you’re weighing lays out what you’d actually be choosing between, in your own words.

If you already keep a journal in Day One or another app, you can bring it with you: export it (JSON or plain text), drop the file in the agent’s files folder and say journal import and the file’s name. Every entry keeps its original date — your 2019 self stays in 2019 — and imported entries live under exactly the same privacy rules as anything you write here.

No, and that is deliberate. It will not say you seem anxious, score your week out of ten, or chart your mood. Those things are easy to build, look impressive in a demo, and are worse than useless when a program is wrong about a person.

Decide is the same: it lays out the choice and then stops. It will never tell you which way to go. It is your life and you are the one who lives in the result.

If what you write suggests you might be about to harm yourself, it saves it like anything else — your words are never rejected — and then says plainly that a program is not the right thing to talk to, and gives you real numbers. Samaritans on 116 123, free, any hour. It will not try to counsel you, because it can’t.

Stricter than anything else in the product. It lives in its own file, never in the memory file, so it cannot end up in a summary or a morning brief by accident. The part that learns your habits never sees an entry. It is owner-only at every level of trust — a trusted person gets your shared calendar, and nobody gets this.

Delete journal overwrites it with nothing. The empty file stays, and that is worth explaining: there is no code anywhere in this product that deletes a file, which is a promise you can check in one search and which the whole thing rests on. Adding one deletion for a good reason is how a promise like that stops being true.

The one exception, stated plainly: reflect and decide send those entries to whichever AI provider you chose, under their terms. That is why they are separate commands you type rather than something that happens to everything you write. Choose Ollama and even that stays on your hardware.

Everything — because there is nothing to move. Your agent was never in your phone: it runs on a computer you leave on, and your phone is just one of its doors. Sign into WhatsApp or Telegram on the new phone and your agent is simply there, with the same memory, the same open loops, mid-conversation if you like. iOS to Android, Android to Huawei, either to a Nokia over SMS — the agent doesn’t know or care which door you walked through today.

Reinstall the app in two taps from the new phone’s browser and you’re done. This is the part phone-makers can’t offer you: their assistants die at the platform border. Yours doesn’t have one.

Copy one folder. Everything your agent is — memory, journal, spending, fitness, reminders, the learning model — lives in agent-data next to agent.js, and your keys live in agent.config.json. Copy the whole folder to the new machine (any OS to any OS — a Windows folder works on a Mac works on Linux, because everything inside is plain JSON and CSV), install Node, run node agent.js, and it wakes up as itself: same memory, same everything.

Say EXPORT in chat any time for a copy of the memory file — a good habit before a move, and proof that nothing about you is trapped in a format only we can read.

Two different questions, and the answers are very different.

To host your agent, a machine needs Node 18 or newer. That covers every current Linux, macOS, Windows, ChromeOS (through its built-in Linux environment), the BSDs, illumos and Solaris — and, genuinely, AIX and IBM z/OS, because IBM ships Node for both. It will not run on MS-DOS, CP/M, OS/360, Multics, TOPS-20, AmigaOS, RISC OS, Plan 9 or a hobby kernel, and nobody should tell you otherwise.

To use it, a device needs to be able to send a message. That is a far longer list, because the agent isn’t running on that device at all — it is running on the one computer you left switched on. A machine old enough to predate the web can reach it by email; anything with a SIM can reach it by SMS.

Will it run smoothly? You don’t have to take our word for it: the agent ships with its own test suite — over 480 automated checks — and you can run every one of them on your own machine with node agent.js --selftest before you trust it with anything. If your computer passes, it runs. If something’s off, node agent.js --doctor tells you plainly which piece is wrong.

Look your system up — every one we’ve been asked about, with a straight answer for each ›

In files on your own machine, in a folder you can open, read and delete yourself. There is no account, no server of ours, and nothing to log into.

The honest exception, stated plainly: if you point it at an AI provider, the messages you send it go to that provider under their terms. That’s true of every AI assistant. Choose Ollama in the config and even that stays on your own hardware.

No. No analytics, no cookies, no pixels, no fingerprinting, on any page including this one. That isn’t a policy we’re promising to keep — there is simply no such code in the site, which you can verify by viewing source.

The only network requests any page makes are for live exchange rates and your own published spreadsheet, both on the money dashboard, and both only if you set them up.

Tell it to. Say forget that about the last thing, or name what you want gone. It removes it from the memory file and from what the learning model has drawn from it — not just from the conversation.

You can also go further and stop it learning something in the first place: say keep anything about Mum private and that subject is walled off before it’s ever recorded. Private list shows you everything currently walled off, and watch Mum undoes it.

It’s a blocklist that runs before anything is written down. Anything matching a private term never enters the memory file, never reaches the learning model, and never appears in a summary — including summaries meant for you.

What it can’t do: it isn’t mind-reading. It matches what you named and close variants of it. If you tell it to keep "Mum" private and then talk about "my mother" without ever connecting the two, it won’t know they’re the same person. Name both and it will.

Only if you switch it on, and it’s off until you do. Say location on to enable it and location off to stop, at any time.

When it is on, it keeps one point — roughly where you are now — and no trail. It does not build a history of where you’ve been, because it doesn’t need one to be useful and keeping one would be the single most dangerous file on your machine.

It’s in a config file on your machine and is sent to that one provider and nowhere else. If you switch providers — from Anthropic to Ollama, say — the old key is not carried over to the new one. That sounds obvious; it is a real bug that shipped in other products, so it’s tested for here explicitly.

Yes. Say someone else is using my phone and it immediately locks your email, files and money, and starts a separate memory for whoever is holding it, so nothing mixes.

Say this is Maya to hand it to a named person. If you’ve trusted Maya, she also gets the shared calendar — never your private accounts.

With your passcode. Set one with set passcode — until you do, handing the phone over still protects your private things, but coming back is just a matter of asking, which is fine at home and not enough anywhere else.

Because you told it. That’s the honest limit and it’s worth stating clearly: it does not read faces or fingerprints, and it never pretends to. If someone picks up your unlocked phone and you haven’t said anything, it will answer them as you.

What it protects against is the ordinary case — handing your phone to someone deliberately — which is the case that actually happens.

Yes. A named person you haven’t trusted gets their own separate memory and no access to your email, files, money or private calendar. Trusting someone is a deliberate act, not a default, and you can undo it.

Hold the microphone button and talk, or hold the M key anywhere on the page. Let go and it sends. What you said appears in the box first so you can fix a mishearing before it goes.

Because a wake word built on browser speech recognition means streaming everything your microphone hears to a server, permanently, on the chance you might say a magic word. Holding a button is more private, more reliable and kinder to your battery.

If you installed the app to your home screen, that’s Apple’s restriction, not a bug here — they block microphone access inside installed web apps. The app detects this and tells you rather than leaving you holding a dead button.

Two things that do work: open the same page in Safari rather than the installed app, or just type. Typing does everything voice does.

Yes, using your device’s own voices — nothing is sent anywhere to be synthesised. It stops talking the moment you switch tabs, and pressing Escape stops it immediately.

Paste the address. Say connect followed by the web address of the thing you want it to watch, and it looks for a door it can use — a feed, a calendar file, a public API. If it finds one it connects itself.

If it can’t find one it says so plainly rather than half-connecting and failing quietly later.

Google, Apple and Outlook, plus anything else that publishes a standard calendar file. That covers most work calendars, school calendars, sports fixtures and bin collections.

No, and it’s worth being precise, because the honest answer is still a good one. It works with anything that has an open door, and there are five kinds: a feed — RSS, Atom or JSON Feed; a published calendar; a public API; email; or, failing all of those, the list of pages a site publishes for search engines. That is a very large number of services — shops, order histories, subscriptions, storage, notes, project tools, release calendars.

The last one is the interesting one, because almost every site has it even when it has nothing else. Point it at a site with no feed and it will watch the page list instead, which tells it when something new appears — but not what that thing says. It tells you that limit rather than quietly pretending the connection is better than it is.

It does not work with an app that has deliberately closed every door and exists only inside its own walled garden. Nothing does, and anyone telling you otherwise is describing screen-scraping that breaks the first time the page changes.

It builds its own connections, so it can reach services nobody made a ready-made link for. It also handles voice notes and photos, runs on the official WhatsApp Business API, and repairs itself when something breaks.

The lower tiers are agents. The Meta-Builder builds agents.

Yes. Claude, OpenAI, Google Gemini, Azure OpenAI, Groq, Mistral, DeepSeek, Grok (xAI), OpenRouter, Together AI, Ollama on your own machine — or anything else that speaks the OpenAI format. One line in the config file changes it; nothing else needs touching, and the named providers don’t even need a URL looked up — the agent knows their addresses.

Your keys never cross wires: pick a non-Anthropic provider without giving it a key and the agent sends no key, rather than falling back to your Claude one.

Three numbers without opening a spreadsheet: what you’ve earned all time, what the business has spent, and what’s left. It reads a sales log you control, converts currencies, and works on any device. There is deliberately no revenue target and no progress bar — a target glued to a screen you check every day turns into a nag, and the true number is in, out, kept.

You can record costs right on it, and keep a customer book — who has which tier, who you gifted it to, who got a discount. Both live in your browser’s own storage (there is no server to send them to — that’s the point), so use the CSV buttons now and then to keep a copy. The actual discount codes and free access happen in your payment provider; the book remembers who and why.

It’s owner-only: set a passcode, add a PIN as a second step if you like, and when your own agent hosts the page it tells you in your messenger the moment the CEO door is opened — if that wasn’t you, you’ll know in seconds.

Yes. Write to it in Italian and it answers in Italian; the model behind it reads and writes most of the world’s languages, so there is nothing to install and no list to check. If it ever guesses wrong — or you share the phone with someone who reads a different language than you write — say speak Italian (or any language) and every reply is pinned to it until you say speak auto.

One honest note: the fixed privacy commands — STOP, WHAT DO YOU KNOW, FORGET — stay in English, and it understands them from anyone. The website you’re reading is in English too; the agent is the polyglot.

Say wake me at 7 or remind me to take the tablets every day at 9 and it will — including genuinely daily ones that come back tomorrow without being asked again. Here’s the honest mechanics: what arrives is a message in your messenger at that time. For actually being woken up, your phone’s own alarm app is louder, works on silent, and doesn’t depend on a computer being awake — the agent says so itself when you set a wake-up, because letting you oversleep on its confidence would be a rotten trade.

For medication and routines, a daily reminder in the app you already watch all day is exactly the right tool — and it’s private to you, like everything else.

It keeps a workout log, and it is deliberately a log rather than a coach. Say workout ran 5k as it happens; ask workouts and you get counts — this week, this month, all time — and your own recent entries back. No scores, no streaks to break, no “you’re behind”: those sell fitness apps and quietly wreck people, and a bad week here just means fewer lines in a file. The log is one plain CSV on your machine, owner-only at every trust level, and never reaches the part that learns you.

Your existing fitness apps link the same honest way everything else does — through the doors they actually have, whatever the OS. Export from the app (Strava: Settings → My Account → Download; Apple Health: profile → Export All Health Data; most trackers have a CSV export), drop the file in the agent’s files folder, and say workout import and the file’s name. It reads Strava-style CSVs, any CSV with date and activity columns, GPX tracks from any GPS app, and Apple Health’s XML — streamed, because those exports run to hundreds of megabytes. Every activity keeps its original date, re-importing doubles nothing, and it reports exactly what it kept and what it skipped. A gym’s class schedule is usually a calendar link it can watch directly. What no app on any OS allows is silently reading your live health data from another app — so nothing here claims to.

Health beyond that is handled with the same honesty: daily medication reminders (see alarms above), appointment tracking through any calendar you connect — and no medical advice, ever, because a program giving it would be wrong at the worst possible moment.

Yes — the dashboard knows what came in, and this is the other half. Say spent £12.50 on lunch as it happens and it becomes one line in a plain CSV on your machine, readable in any spreadsheet. Spending gives you the month’s total and top categories; budget £200 a month on eating out sets a line it watches. You hear about a budget at 80% and again if you go over — and that is all that happens. It counts; you decide. It will never moralise about a purchase or refuse to record one.

What you spend is treated as a record, not a personality: spends never reach the memory file, the part that learns your habits, or your preferences — and like everything with money in it, they’re owner-only at every trust level and locked when the phone is handed to someone else.

Because reading a live bank account is Open Banking — a regulated activity with an FCA licence behind it, and this product doesn’t have one. Most apps in this position rent access through a licensed middleman and pass your bank credentials through a company you’ve never heard of. We’d rather tell you the honest version.

The honest version works fine: almost every bank and money app — Monzo, Starling, Lloyds, Barclays, Revolut, YNAB — exports CSV. Download it, drop it in the agent’s files folder, say spending import and the file’s name. The importer copes with real exports (dates written backwards, descriptions full of commas, separate money-in and money-out columns) and then reports exactly what it did: spends kept, money-in left out, rows it couldn’t read, rows it had already seen. Re-importing the same file twice doubles nothing. The truth arrives a day later and belongs entirely to you.

No. Publishing a Google Sheet is the easiest route, but agent.js --serve will host the sales feed itself if you’d rather not involve Google at all.

Set owner_passcode and the dashboard asks for it before showing anything.

Be clear on what that is and isn’t: it hides the numbers from someone who wanders onto the page. It is not encryption. The published spreadsheet link is the real secret — keep that private and don’t share your dashboard URL.

The pages you’ve already visited do, including the dashboard with the last figures it saw. New sales data obviously needs a connection to arrive.

Run node agent.js --doctor. It checks each connection in turn — your AI provider, your messenger, your calendar, your sales feed — and reports which one failed and why, in plain English rather than a stack trace.

Nine times in ten it’s a key that was pasted with a space on the end.

Node isn’t installed. It’s free, takes about two minutes, and comes from nodejs.org — choose the version marked LTS. Then run node --version to confirm, and you should see a number of 18 or higher.

Start it with node agent.js --guard instead. That runs a small supervisor which restarts the agent if it dies and recovers its state, so a crash at 3am doesn’t cost you the morning.

Correct it in the conversation and it updates. If it’s persistently wrong about a pattern rather than a fact, ask what have you learned — you’ll see what it has concluded and can tell it to drop the specific thing.

Almost always the AI provider rather than the agent. A smaller model on the same provider is usually several times faster and, for routing and short replies, no worse. The config lets you set a different model for quick decisions than for real thinking.

Say so, with the sentence you’d say to a friend — "it stopped replying after I connected my calendar" is more useful than a log file. Every report gets read.

Two things, and neither is a fight. First, the bug gets fixed: the agent repairs most faults itself (a self-healing layer, a watchdog that restarts a dead process, a redundant copy of everything important), and whatever it can’t fix, you report and it’s fixed in an update you keep. Second, a genuine product fault that cost you time earns you three months free on top of your plan — on Forever, the cash equivalent: a partial refund of three months at your tier’s quarterly price. Report what happened and roughly when, and that’s the whole procedure.

This sits on top of the 7-day promise and your statutory rights, not instead of them. Honest edges: it covers faults in the product, not your AI provider’s outage or your messaging platform changing its rules.

Seven days, measured to the minute: 168 hours from the purchase time printed on your receipt, so buying at 2:32pm gives you until 2:32pm a week later — and it’s when your email is sent that counts, not when it’s read. Ask inside that window and you get your money back — no form, no reason needed, and no attempt to talk you out of it. One line by email is enough.

It’s once per person: the promise covers your first purchase. Refund and buy again later and the second purchase is final — you knew exactly what you were buying the second time. Your rights if the product is faulty (and the make-good promise) apply to every purchase regardless.

One thing to know before you buy: because delivery is instant, checkout asks you to agree to immediate supply and to downloading ending your statutory 14-day cooling-off right — that’s the standard trade UK law sets for digital downloads. The 7-day promise is ours, sits on top, and isn’t affected. Nor are your rights if the product is faulty.

Nine rungs, so nobody pays for room they don’t need: Solo (just you, £9 a quarter), Duo (the two of you, £14), Personal (5 people), Smart Home (10), Team Hub (15), Meta-Builder (20, and it builds its own connections), Estate (Meta-Builder for 35 — the same as buying 15 extra seats, minus £5 and the admin), and Concierge (Meta-Builder for 50, the biggest fixed room — honestly: Estate plus 15 more seats, £5 more for one bill) — and for teams of 100 or more, the Custom card prices per person at checkout. The tiers differ in what the agent can do. The billing choice is separate: any tier can be paid monthly, paid quarterly, or bought outright — all three are the identical product. Fair warning we also print on the pricing page: monthly works out about 20–25% dearer than quarterly; it exists for people who prefer a smaller number leaving their account each time, not because it’s a better deal. Whichever you pick, only the payment schedule differs.

Yes — on Meta-Builder, Estate and Concierge, at any time, on any billing plan. On the quarterly plan an extra person is £5 a quarter. On Forever it follows the same fifteen-year formula as everything else: £300 once, and that person is part of your plan for good — bought a year after the plan itself if you like. The “add people” link on the pricing page sells any number in one checkout.

The other tiers have fixed rooms — Personal 5, Smart Home 10, Team Hub 15. Need more than your tier holds? Moving up a tier is the honest answer, and past twenty people, Meta-Builder grows a seat at a time.

The honest mechanics, since there are no accounts here: the payment is the licence, and adding the person to your agent is one message — “this is Maya”, or a seat in the config. Your agent was always able to know your household; the purchase makes it yours to use for more of them.

One payment, priced at fifteen years of that plan. After it there is nothing to renew and nothing to cancel.

Worth doing the arithmetic yourself before choosing it: fifteen years is a long time to bet on any piece of software, and the quarterly plan gets you the same product. It exists for people who genuinely prefer never thinking about a subscription again.

Your AI provider charges for what you use — typically a few pounds a month for personal use, and nothing at all if you run Ollama on your own machine. That money goes to them, not to us, and you can see it on their dashboard.

Nothing else. No hosting, no per-seat fee, no upsell inside the product.

Fixes for genuine product faults are a promise — see “What happens if something breaks?” above. Beyond that, updates are a bonus rather than a commitment — and when they happen, they’re files you keep, not a service that can be withdrawn. Because everything runs on your own machine, a version that works today keeps working whatever happens to us — which is the point of building it this way.