What changed.

The little things we’ve added and made better. Newest first.

What changed

One block per release, newest first, in plain words. Your box updates from the signed release once you say yes on its card; Settings shows which release it is on. This file is the source of truth: the public page at https://mymagicbox.lovable.app/changelog is this file, read from the kernel's release branch. A release cannot be signed without its block here (bin/release.mjs checks). The nerd detail behind each line is in docs/BUGS.md, by row number.

Release 40 · 21 Oct 2026

Every page has its own chat. Pages → a page → ⋯ → Chat. Your box's page builder now talks about a page only there: its questions, the card, "v2 is ready", what failed. It no longer drops those into whatever chat you have open; the chat you asked from gets one line pointing you to the page's chat. Ask for a feature or a fix in a page's chat and it goes to that page; you don't have to name it.
"Use my Claude subscription" now keeps working after you sign in. Claude's sign-in program sends its token in pieces, and your box sometimes saved only the first piece, so it looked signed in but every answer failed. It now waits for the whole token. If you signed in before, sign in once more.

Release 39 · 21 Oct 2026

Connecting the Laso wallet with a key works. Your box now hands Laso the key the way Laso expects, and you can paste the key on its own, the "Authorization: Bearer …" line from Laso's dashboard, or Laso's whole "paste this to your agent" message: your box finds the key in it.

Release 38 · 20 Oct 2026

Dark mode is readable. If your phone is set to dark, every screen of your box now has proper dark colours: text, cards, chat bubbles, buttons, the menu, the creature, the Memory cards and the locked screen are all easy to read. Nothing changes in light mode.
"Use my Claude subscription" no longer gets stuck signing in after you paste the code. If Claude refuses the code, your box now says so straight away; if Claude doesn't answer, it gives up after a minute and a half and lets you start again.
The chat box grows as you type, up to about half the screen, then scrolls. It shrinks back after you send. Shift+Enter starts a new line.
The nightly read also makes to-dos and asks you about anything unclear. Things you agreed or were asked to do become to-dos (at most 5 a night), each saying where it came from; one you've dropped never comes back. When something matters and your box can't tell (is the dinner still on, which date is right), it asks you, at most 2 a night, and shows where the question came from. Your answer goes into your memory in your own words.
Laso wallet (off until you switch it on). Settings → Laso wallet, with your passkey. Connect your Laso account (or make one on Laso's page), top it up by sending USDC on Solana or Base to the addresses it shows (with a QR), then just ask in chat: "Venmo Jo $40 for dinner", "get Mum a $25 Starbucks gift card", "make me a $5 card for my coffee". Every payment is a card that shows the fee and the total, and nothing is paid until you approve it with your passkey, whatever your other switches say. Your box keeps its own caps ($50 a payment and $150 a day to start; raising them needs your passkey) under Laso's. A card's number or a gift card's code shows only in Settings, on your passkey; your box's brain never sees them. There's no test mode at Laso: every payment is real money.

Release 37 · 20 Oct 2026

Learning from your day is pickier, after its first real night. Someone who only booked one meeting with you no longer gets a page of their own. A weekly event is remembered once as a habit ("lifts on Saturdays at 12:30"), not once per week. It uses your own words for your people (your wife stays your wife), and it no longer writes facts about "an unknown contact".

Release 36 · 20 Oct 2026

"Use my Claude subscription" no longer gets stuck on "Getting a sign-in link from Claude…". Claude moved its sign-in page to claude.com and your box was still looking for the old one.
Pages that Claude or ChatGPT write go through the page factory too. Your box writes the checks the page must pass, asks you its questions, tests their page with your own notes and fixes what fails before it's under Pages. Nothing they send is installed as sent, however they send it. Nerds: BUGS 212.
Your box learns from your day. Every night before it tidies its memory, it reads what came in and went out since it last looked: your email (every Google account), WhatsApp, Telegram, iMessage and your calendar. It writes down what's worth remembering (who people are to you, plans and promises, birthdays, what you like) and files it into your memory. It skips newsletters, adverts, sign-in codes, links and group-chat chatter, and it never keeps the messages themselves, only what it learned. Last night's card says what it read and learned, and every fact shows where it came from, with "I didn't mean that" to remove it. Switch each account on or off in Settings → Learn from my day, or forget everything it learned from one. It stops at $0.25 of thinking a night. Nerds: BUGS 213.
When you connect something, it looks at the last month straight away, then tells you in chat what it learned and asks a few questions so it gets your people right ("Who is +1 512 555 0123? You message them most days").
One small question a day, to get to know you ("Morning person or night owl?"). Never more than one, never a notification, always skippable; answer on Home or in chat. Switch it off in Settings → Learn from my day.
Your people, matched to their numbers. A person in your memory can carry their WhatsApp, Telegram, phone and email, so messages read "Jo (your sister)" instead of a number. The box only links a number to someone when it's sure; otherwise it asks you. Numbers never go to the brain, and assistants you connect don't see them. Nerds: BUGS 214.
The old "Morning catch-up at 7:00" line in what your box knows about you is gone, since the morning paper is off.

Release 35 · 20 Oct 2026 · UI overhaul

UI overhaul. Your box has a new face and a new shape. No menu bar across the top: a round button at the bottom opens a menu that slides up, with Home, Chat, Memory, Cards, Connectors, Pages and Settings. Home is your box as a creature, with its mood on its face and one line about how it's doing; tap it when it needs you and the cards open. Cards are one line each with Yes and No, and unfold on a tap. Chat fills the screen and its messages scroll on their own. Memory opens on what happened last night, what your box knows about you, your maps, who uses your memory, and the things it set aside. Pages shows every page with its version and checks, what's being made right now, the questions your box has for you, and every older version to go back to. Settings and Connectors keep everything they had, in the new look. It's the same design you can try at https://mymagicbox.lovable.app/preview, now running on your own box with your own memory.
Your memory no longer loses things. A line the nightly tidy couldn't place, or thought was small talk, used to vanish after three tries. Now it waits in Memory → Set aside with the reason, in full, and you can file it again or throw it away. A line you send back is kept for sure. The tidy also reads long lines whole, splits a line with several facts across the right notes, starts a new note when none fits, and keeps how you like to work. Nerds: BUGS 210.
Chats come in sections: Home, Money, Family, Work, Health, or ones you make ("make a section for the house hunt"). After a couple of messages your box names the chat, puts it in a section and adds a few tags. Anything you set by hand stays as you set it.
The morning paper and evening edition are off. Your box was in your face too much. The nightly memory tidy still runs at 3:00, and now you can see what it did: "Last night I filed 14 things into 6 notes and started 2 new ones (House hunt, Photography)." You can switch the paper back on in Rhythms.
Cards fold to one line ("Send a WhatsApp to Sam", "Update to release 35"), with the full text one tap away. What you approve with your passkey is still the full text.
Your box has a mood: happy, sleepy, thinking, needs you, hungry or not well, with one line saying why. Nerds: BUGS 211.
Pages take their time now, like a real software team. Ask for a page and your box first writes what it's for and the checks it must pass, has those checks reviewed, and asks you a question or two if something is unclear. Then it designs the page, builds it, tests it against every check with your own notes, fixes what fails, and only then puts it under Pages, telling you how many checks pass. Each job is capped at $1.00 of brain spend and 20 minutes.
Every change to a page is a new version that keeps what it was for, its checks and what changed. The old checks run again, so a change can't quietly break what already worked. You can go back to any of the last 10 versions, and compare two.
New pages live under Pages instead of filling up the top menu. Pin one if you want it in the menu.
Claude and ChatGPT can ask your box to make a page the same way (request_page), and see its questions; what they send is always a card.

Release 34 · 19 Oct 2026

Telegram can be linked from Connectors → Telegram: type your number, then the code Telegram sends you (and your Telegram password, if you have one). Your box joins as one of your devices. It reads your chats, sends a message only after you say yes on a card, and only takes orders from your own Saved Messages, when you start with its name. Everyone else's messages are read, never obeyed. Reminders and routine results land in your Saved Messages too.
Telegram needs the maker's Telegram app key. New boxes get it automatically; an older box picks it up on its next check-in with the website, and only while it has none. Nerds: BUGS 209.

Release 33 · 19 Oct 2026

Send your box a link and it reads it: an article, a recipe, a listing, a menu. It no longer has to build a connector first just to open one page. It only opens links you send it yourself, never one it found in an email or on another page, and it tells you plainly when a site blocks it, needs a login or only works in a browser. Nerds: BUGS 208.

Release 32 · 19 Oct 2026

The screens catch up with release 30. Settings now shows the five switches (What I may do), Lock my box and your recovery phrase, Updates (Ask me), Who can see my box, Settings → Brain with your own key or your Claude or ChatGPT plan, and every Google account you've added, each with Reconnect, Add Drive, Make main and Unplug.
Cards have short titles (New page, Send as you, Save this?…), a "Sent by Claude" chip when another assistant sent one, and a Done label with Undo for things your box just did.
Connecting Claude or ChatGPT: three checkboxes instead of four choices: Memory, My messages and mail, Can build for my box.
The Skills tab is now Routines; starter pages from the library are under Pages → Add from the library.
Update cards show what changes, with Update now, Not now and Skip this release.
Removing the last passkey that opens your lock asks you first ("I have my recovery phrase"). Today shows how many cards are waiting, and when Today is hidden or replaced, a chip in the header still takes you to them.
Flipping Safe mode, or any switch under Things I built, changes at once and no longer reloads the page or closes the section you're in. Nerds: BUGS 207.

Release 31 · 19 Oct 2026

Claude and ChatGPT find their way round your box better: they're told what they may do, how your memory is laid out, how to build a page, and exactly what to tell you when something needs you (a box to tick, an account to connect, a locked box). Searching your memory for your own name now finds you.
The new screens for release 30 were meant to ship here but weren't included; they come in release 32.

Release 30 · 19 Oct 2026

Your box can lock itself so only you can open it. Its memory, accounts and files move onto a locked part of its disk, opened only by your passkey or a 12-word recovery phrase you see once. The people who make magic box can't open it, and neither can Google. A new box asks right after you set your passkey; an older box locks when you tap Lock my box in Settings → Safety.
The moment you first lock your box, the makers' ways into the machine close: no remote login, no scripts from the cloud console, no serial console. Until you lock it they stay open, and its disk can be read by whoever runs the machine, so lock it.
After the computer itself restarts, a locked box waits at a locked page until you open it with your passkey, and tells you when it restarted. Restarting your box from Settings, updates and crashes don't lock it, so the morning paper and the nightly tidy keep running. Claude or ChatGPT are told plainly that your box is locked.
Updates wait for your yes. This is the last one that installs on its own: from now on each new version comes as a card, Update to release N, that shows what changes. Your box keeps doing its jobs whether you say yes, no or nothing.
The website can't change your box's Google sign-in any more, and it now hands your box its keys once, when the box starts, sealed on the box instead of written into the machine's setup.
Several Google accounts on one box: add your work account beside your personal one (Connectors → Google → Add another Google account). Your box reads mail, calendars and Drive from all of them, says which account each thing came from, and puts a draft in the right one. Unplug one, or pick the main one, next to it.
Bring your own brain. In Settings → Brain, paste your own Venice, Anthropic, OpenAI or OpenRouter key and pick any model it offers, or sign in with your own Claude or ChatGPT subscription. It's your own bill, so the monthly budget doesn't cap it. If it fails, your box says so, and the included brain answers instead unless you switch that off. Your key is sealed on the box and never shown again.
Five switches, and they all really work. Settings → What I may do: Change my box (Just do it, with Undo), Send as me, Share outside my box, Give other assistants access, Delete for good. No more modes. Ask for a page, a look or a routine and it just happens; sending, sharing, giving access and deleting still ask first. After your box has read an email, a file or a web page, anything that sends, shares or deletes is always a card.
One word for your own pages: Pages. The Skills tab is called Routines, and ready-made pages from the library are starter pages, under Pages. Cards have nine plain titles, and a card another assistant sent says who sent it.
Keys for other assistants have three checkboxes: Memory, My messages and mail, Can build for my box. Tick any of them; reading your memory is always in.
The manual your other assistants read (/for-bots) now starts with how to connect, every tool in one line, and how to build you a page end to end, and says how to use your box as one shared memory.
Some of the screens for this (Lock my box, Brain, the list of Google accounts, the switches and the three checkboxes, the Routines name) arrive with the next release of the box's app; the box's side is all in this release.
Nerds: BUGS rows 125 to 140; docs/SITE-LOCK.md, docs/UI-LOCK.md, docs/UI-BRAIN.md, docs/UI-MULTI-GOOGLE.md, docs/UI-SIMPLE.md.

Release 29 · 19 Oct 2026

Your screen is yours to arrange. Ask your box for a hamburger menu, fewer tabs, a different order or different names, what opens first, or one of your pages as a tab or as the home screen. It becomes a card (or just happens, if you set "Arrange my screen" to Just do it in Settings → What I may do), and undo puts it back. Settings always stays. Claude or ChatGPT can send you an arrangement too. The screen itself follows in the next release of the box's app.
A page can be your menu, your hub or your whole front screen: it can now open your other pages and tabs, read and tick your to-do list, see which cards are waiting for you, and read the diary.
Your box can build itself a connector that reaches any website, when you say so on its card. That is how it reads a link, a video, a repo or a tweet you send it: it builds the reader, you approve it once.
The box's standing order has changed: when you ask for something it can't do, it builds it (a page, a connector, a layout change, a look change) instead of saying no.
Nerds: BUGS rows 122 to 124.

Release 28 · 19 Oct 2026

Pages no longer add tabs to the top menu. The box keeps its seven tabs; every page is listed under Pages and opens in place with a back button. A page that is off still shows there with its status, and opening it takes you to Settings → Things I built. From Rava's owner report.
The icon you asked for in chat now shows by the box's name and as the browser tab icon, on real boxes and in the demo.

Release 27 · 19 Oct 2026

Pages work on iPhones. Safari treated a page's calls to the box as forbidden, so no page ever showed data or saved anything there; the box now names its own address in the rule that allows those calls. A page you already built needs one more open.
The chat window has a fixed height with its own scroll bar, so the box to type in stays at the bottom and the page no longer grows forever. New messages scroll into view unless you had scrolled up. The first fix from an owner's "Tell the maker" report.
Ask for a different icon ("make your icon a little box") and the box changes the mark by its name and the browser tab icon.
Pages live under the Pages tab; the box says "under Pages" instead of "in your menu".
WhatsApp linking by code is confirmed working on a phone.

Release 26 · 19 Oct 2026

Settings → Other assistants offers the new kind of key, "Can read my messages and mail", and shows it in the connections list.
Things I built and the Pages tab say when a page reads all your memory and which accounts it uses.

Release 25 · 19 Oct 2026

"Remember that…" just remembers. When you ask in your own words to remember, note, save, add to your list or remind you, the box writes it, even in a turn that read email or messages. A card still comes only when a fact surfaced out of outside content without you asking.
Your other assistants can read your messages and mail. A new choice when connecting Claude or ChatGPT, or a new kind of key in Settings: "Read and write, and read my messages and mail". That connection can read your WhatsApp, iMessages, email and calendar through the box, never send, and every read is in your diary. Claude, ChatGPT or an assistant on your own computer only.
A page can see all of your memory when its card says so, and it can change or forget facts in the notes it reads. Your Undo brings any line back.
A page can use your accounts when its card lists them: it reads your WhatsApp, iMessages, mail, calendar or Drive through the box. Sending from a page is always a card.
Nerds: BUGS rows 116 to 118.

Release 24 · 19 Oct 2026

This page. Every release from now on says what changed, here and at /changelog on the site, and Settings → About on your box links to it.
A release cannot be signed without its entry here.

Release 23 · 19 Oct 2026

Apps and sections are one thing now: a page. Every page lives in your menu, keeps its own notes in your memory and reads the maps you allow, so a page and your chat know the same things. Pages you built before were moved over with their data. The Apps tab is now Pages; Things I built lists Connectors, Pages and Looks.
A new page is in your menu the moment it is built. The Activate step is gone. If a page breaks the first time it opens, the box takes it out and says so.
If a page opens but never talks to the box, the diary now says so and names the cause.
WhatsApp linking: one code per attempt, and asking again shows the same code for about ten minutes. The sheet counts down. A half-finished attempt can no longer spoil the next one, and the diary says when your phone accepted the code, when it ran out, and why an attempt stopped.
Every page remembers who made each version.
Pages have a documented address (/s/your-page/ on your box, your own sign-in needed).
A connector that never leaves the box no longer has to name a website.
When the nightly tidy drops a line it couldn't file, the diary names it and why.
The sandbox's own security warning is gone from connector test logs.
From Lili's Claude's report: docs now say the sign-in limit is for connectors builders write, not the box's own Google connector.
Nerds: BUGS rows 105 to 115.

Release 22 · 18 Oct 2026

Chat threads for everyone: a thread bar with General and + new, and each thread can be renamed, cleared or deleted. Memory is never touched by any of those.
Tell the maker: a feature request, a bug or a question, from chat or from Settings. You see the exact report first; it carries your words and the box's own status lines, never your memory, chat, messages or mail. The makers reply by email.
The WhatsApp sheet leads with the laptop route: open your box on a laptop or tablet, show a QR, scan it with your phone.
Nerds: BUGS rows 101 to 104.

Release 21 · 18 Oct 2026

Google connect no longer ends on "bad_origin".
The iMessage helper installs cleanly (the Mac's security tool was being sent a command it doesn't have). A fresh install line and code are needed.
WhatsApp: the box now announces itself as a Mac companion, which WhatsApp accepts; a failed link writes WhatsApp's own reason in the diary.
Nerds: BUGS rows 98 to 100.

Release 20 · 18 Oct 2026

No more invite codes. Asking for a box carries a private note (who you are, what you want help with) that only the makers read.
Private-beta copy wherever someone asks for a box; setup says "picked", never "connected", until a box exists.
Voice notes are described as what they are.
Nerds: BUGS row 97.

Release 19 · 18 Oct 2026

Updates that need something installed on the machine now do it themselves, so no feature ever needs a Rebuild (Rebuild erases a box; Restart keeps everything).
Nerds: BUGS row 96.

Release 18 · 18 Oct 2026

The box builds itself: connectors in a sandbox that reach only the websites on their card, sections (now pages) with their own notes, and changes to how it looks. Every one starts as a card you approve with your passkey.
Your other assistants can build for it: a key that "Can propose changes to your box" lets Claude, ChatGPT or Codex send a connector, a page or a look change; each arrives as a card that says who sent it.
Google Drive, read-only, as an add-on to Google. iMessage from your Mac with a small helper.
Three security reviews fixed; Google sign-in is bound to the device that started it.