Browse the docs

What’s new

What changed in each release of Archie

The notes Archie shows when it updates, newest first. Archie updates itself when it opens, and you can press Check for updates in Settings.

On this page

Archie 0.3.12

October 26, 2026

  • Your agent can work numbers out in a short Python program, on a Mac, instead of in its head: a total, a percentage, a loan payment, the date 45 days after an invoice, or the figures in a spreadsheet you gave it. The program runs inside Archie with no internet, your files are never changed, and How I worked it out under the answer shows you the program. Python is a one-time download of about 23 MB that Archie fetches on its own.
  • Show your agent a website task once. With Computer control on, press Show me once on its panel and do the task in Archie's own browser window, talking it through if you like. Your agent keeps the steps, and next time it does the task the way you showed it. It never keeps what you typed.
  • An empty chat suggests two or three things to ask, picked from the skills your agent has and the accounts you connected. Tap one to send it.
  • A line under a reply names a note your agent saved to memory, with Open and Undo.
  • A guide for moving your agent to a computer that stays home, so it keeps working when your laptop is closed.
  • From Archie Mobile you can now let people in, turn them away or block them, switch Call businesses on or off, change the email and text watches, and say whose each mail account is.
  • Archie tells you when your chat app stops reaching your agent, in Archie and on Archie Mobile.
  • A file opened from an email comes into the chat, instead of a message naming where it was saved, and files go out as big as each chat app takes.
  • Spoken replies on iMessage arrive as audio messages that play in the thread.
  • The Dashboard's status sits on a card with a sky that follows the time of day.
  • On Archie for Teams, a teammate on Everything you can gets websites in a browser window of their own, which is not signed in to anything.
  • Add-ons that keep dates do their main job again. Bills & Subscriptions rolls a bill forward by its own frequency, Home Maintenance and Plant & Pet Care mark a job done today, and Habit Tracker, Learning Coach, Stay in Touch and Price Watch write today's date when you report something, instead of being refused.
  • Fixed: a text reply waited forever on the Mac's Messages permission prompt. It now says what to press.
  • Fixed: after the Mac woke from sleep, Archie Mobile could take a while to catch up.
  • Your agent is less likely to take back a correct earlier answer by saying it could not confirm it.
  • Every dialog has a visible way to close, and long links wrap instead of making the chat scroll sideways.

Archie 0.3.11

October 19, 2026

  • Bring in what ChatGPT, Claude or Gemini knows about you. Paste its answer on the Memory screen, tick the lines to keep, and your agent starts from them. A line with a password, a card number or a Social Security number is never kept. You can also change any memory line in place.
  • Open Archie from any app with Control, Shift and Space on a Mac, or Ctrl, Alt and Space on Windows. Change the keys in Settings, under General. On a Mac, Archie also has an icon in the menu bar.
  • Calendar invitations with people and a place, on Google and Outlook calendars, after your yes like every calendar change. The card names everyone who will get an invitation.
  • Your agent can see when other people are busy, never what they are doing, when it looks for a time, on Google and on a work or school Outlook account. On Google, reconnect first: on the Connections tab, under Email, calendar, files, press Reconnect on your Google account, keep Let it change events ticked, and sign in with Google again.
  • Stop asking about your own calendar. After you say yes to three changes, the card offers Yes, and stop asking. From then on, adding or moving an event only you are on happens at once, with Undo over the message box. Deleting, inviting and every reply still wait for you.
  • Charts in the chat. Ask for a comparison and your agent can draw bars, a line or a pie.
  • Search this chat with the Search this chat button or Command F (Ctrl F on Windows), and on Archie Mobile.
  • Talk over your agent to stop it on the talk screen, and the chat marks where it stopped.
  • Your agent can phone a business for you, if you want it to. It is off until you switch on Call businesses on your agent's Setup tab, under What it can do, and paste a key from your own Bland account, which bills you by the minute. Your agent shows you a card first, and only your press on Place the call dials. Businesses only, never a person you know.
  • A line over the message box says what changed and what went out today.
  • A line under a reply says what your agent read to answer it.
  • A draft says what it added and where that came from, and offers Write like me on the card.
  • Keep a list matched to its spreadsheet, or to a Google Sheet shared with your agent, from the import screen. Archie checks the file every minute and before the morning list.
  • The morning list puts what needs your decision first, and says where a booking and your calendar disagree.
  • Email Manager and Package Tracker ask before they read your inbox. Nothing new is read until you press Start reading my mail.
  • Ask where your information goes and you get a straight yes or no first.
  • The Routines tab says routines run on this computer, and that one due while it sleeps runs when it wakes.
  • On Archie for Teams, your agent uses a teammate's own phone and email when they are the one asking, tells you what a teammate saved, and lets a teammate answer their own cards from their phone.
  • The daily briefing sends nothing on a day with nothing on the calendar.

Archie 0.3.10

October 17, 2026

  • Google has verified Archie, so connecting Gmail or Google Calendar goes straight to Google's permission page, with no warning page before it.
  • On Archie for Teams, people have roles. Make a teammate an admin, who can change the agent and who is on it, set what each teammate can do, or hand the agent over to a new owner, with both of you told. Archie knows who is at this computer and speaks as them, and an admin can manage the people from their phone.
  • Your agent changes the spreadsheets, Word files, decks and text files you give it in their own style, and reads a spreadsheet's row numbers, column letters and formulas, so a change lands in the right cell. A file it cannot open gets a note saying how to export one it can.
  • Job Hunt can start from the people you know. Bring in your LinkedIn connections file and it shows where they work and names one person a week to write to.
  • A reply shows the message it answers at the top of its bubble.
  • The chat moves more smoothly. A reply types itself out in its own bubble, and the chat glides to the newest message and stops following only when you scroll away.
  • The back arrow sits beside each page's title.
  • Your agent's browser works with a bank's dropdown menus, finds a file it just downloaded by its name, and leaves no blank tabs behind.
  • A skill's checklist stops looking unfinished once every step is done.
  • On the Account page, Archie asks how you heard about it. Answering is optional, and nothing is sent unless you press an answer.
  • Fixed: a Google refusal said Google might be stopping Archie. It is just a refusal now.

Archie 0.3.9

October 14, 2026

  • Archie Mobile shows notifications for reminders, reports, new mail, and replies to what you sent from your phone, even when the app is closed. Turn them on in Archie Mobile, under Settings. What each one says is locked on this computer and opened only on your phone.
  • Close Archie's window and your agents keep working. Archie opens at login with no window, and only quitting Archie switches it off.
  • A routine your computer slept through runs when it wakes, and Archie says when a routine that has never run missed its time.
  • On Archie for Teams, teammates can use Archie Mobile. Add someone with an Archie Mobile link and they join from it, with a conversation of their own, a Dashboard of their own, and the agent's lists, showing their own entries and the ones nobody is named on. A Telegram invite joins with one press too.
  • Ask what the team has been doing and your agent tells you what each teammate asked. You can set a reminder for a teammate, and it reaches them in their own chat.
  • On Archie Mobile, press a list to see what is on it, then add an entry, change one, or delete one. You can also add, change, or remove the people on an agent from your phone.
  • Each morning your agent sends a list of what is on for the day, made for your line of work. After your first questions it says when the first one comes.
  • The first questions are fewer, and a question of yours ends them. When a reply names an app you have not connected, a button under it opens where to connect it.
  • A spreadsheet dropped in the chat or opened from your mail goes on a card that shows its rows before anything is added, and says which columns the list has no place for.
  • Your agent can write a first text to a new lead, and says whether they agreed to texts. Archie copies it for you, and Archie Mobile opens Messages with it typed in.
  • Under a draft, a note names anything in it that nobody told your agent. With sending off, a mail card is a draft to copy, with no Send button.
  • Undo all under a reply takes back everything it saved.
  • Say Monday on a Monday and your agent asks which one you mean. It says when a date it filled in is a guess, and asks before writing over a date you gave it.
  • Real estate deals have inspection, appraisal, and loan approval dates. A lead with the same name, phone, or email as one on your lists is named as already there.
  • Ask what Archie costs or how much free credit is left and the answer is exact. An add-on that needs your own AI account says what one usually costs and has a button to set it up.
  • A request to pass on new or changed wiring instructions is turned down every time, before the AI is asked anything.
  • Pairing Archie Mobile no longer moves an agent off its chat app. Archie Mobile is a box you tick beside it.
  • Fixed: the agent page could grow wider than the window.
  • Fixed: mail waiting to be read could lose its cards if Archie restarted.
  • Fixed: reconnecting your email could lose your signature and your shared calendar.

Archie 0.3.8

October 12, 2026

  • Your agent's first questions start with the first thing you type. The first screen has a message box, and the questions ask what your days look like and where you keep your work. Archie for Teams starts with a pack of its own.
  • Written setups for four lines of work: portable restrooms, restroom trailers for events, home trades, and real estate. Each holds the lists that work runs on and a few rules a new hire would hear on day one, and you can change any of it by asking.
  • Lists keep their stages and sit in your agent's sidebar. The Dashboard shows each list with a bar of where its entries stand, and a list says how many are entered out of how many in all. Ask in the chat and your agent sets up a list's stages itself.
  • Bring a spreadsheet into a list by dropping it in the chat, or with Import on the list itself, on the Knowledge tab, under Lists. Archie names any cells it could not read and never adds the same unit twice.
  • Entries your agent reads off a photo, a page, or your calendar go on one card you check. Press Add and they go on the list as one job, with one Undo.
  • Under each reply, what your agent saved and what it read, with Open and Undo. The top of the Knowledge tab shows what your agent keeps.
  • Your agent keeps your business's name, prices, phone number, and sign-off and uses them in what it writes. When a draft needs a detail nobody gave, it leaves a blank, and an email or text with a blank in it will not send until you fill it in.
  • Ask who owes you money and your agent answers from your lists, the chat, and its notes. Ask who to call and it works that out from your People and Deals lists.
  • Ask what's on, what's due, or what's free on some days, and your agent looks at your lists, calendar, and reminders for those days before it answers. The Calendar shows the dates on your lists, and Archie tells you when a booking puts a trailer or a unit in two places at once, whether your agent saved it or you typed it on the Knowledge tab, under Lists.
  • Every reminder shows on your to-do list, and deleting it in either place takes it off both. When a reminder goes off, Remind me later lets you pick any time, and each button says what it does.
  • Right-click an agent to open it, start or stop it, rename it, change how it looks, or delete it.
  • New chat starts over with one press, and a line marks where each day and each new chat begins.
  • Edit it opens a pop-up that shows the message your draft answers. Rewrite it for me takes more from you, and looks things up when you point it at them.
  • When the AI your agent thinks with runs out, it says so once, with the ways to keep going, and keeps what you send until the AI is back.
  • When the AI company fails, Archie asks it once more before showing anything, says whether anything was saved or sent, and puts Try again under the reply.
  • A reply that says your calendar or email is not connected has a button under it that opens where to connect it.
  • The price is on the trial strip, the AI account panel, and a new Your plan row on the Account page. A plan with no AI in it asks for your AI account right after you buy it.
  • Two new guides: where your information goes, and what Archie connects to.
  • A stopped agent says who stopped it and when.
  • On Archie for Teams, the people you add are teammates. Write down who each one is and what they do, and your agent knows its team. A button beside each teammate shows what they can see and change, and the Dashboard shows what each did this week. A teammate can set and cancel their own reminders, and asking about the work gets them their own entries and the ones nobody is named on.
  • A backup from Archie opens in Archie for Teams, so you can move your agents over when you hire.
  • Your agent is told never to pass on new or changed payment instructions, even when you ask, and to tell you to confirm by phone at a number you already have. It is an instruction to the AI, and an AI can get it wrong, so read anything about moving money yourself.
  • Fixed: a reply could say it saved something to your lists when nothing was saved. A note under it says so now, with a Save it button.
  • Fixed: the buttons on your agent's question cards stopped working when Archie restarted. They last two weeks now.
  • Fixed: a reply could include your agent's notes to itself between searches.
  • Fixed: questions about your own work could go to the web researcher, which reads only the one message it gets. Fewer do now.

Archie 0.3.7

October 9, 2026

  • Archie for Business is now Archie for Teams. Same app, same agents, same plan, and nothing to set up again; only the name changed. On Windows the update takes the old one's place, with your shortcuts and open at login carried over. On a Mac, System Settings may still list it as Archie for Business.
  • Archie for Teams catches up with Archie 0.3.5 and 0.3.6, which it had been waiting on: the backup button on the plan screen, the search that waits after your agent reads new mail, and the rest of both releases.
  • On Archie for Teams, a new choice for the people you let in: Talk and add to your lists. A teammate can add to or change the lists on your agent's Knowledge tab without being handed your mail, calendar or connected services.
  • On an Anthropic account, Economy now runs on Claude Haiku 5.5 and Balanced on Claude Sonnet 5.5. A model you picked yourself stays as it is. Model names now read the way people say them, Claude Sonnet 5.5 rather than claude-sonnet-5-5.
  • Checking off or editing a to-do in Todoist or Google Tasks happens at once, with an undo that puts it back. Deleting one still waits for your yes.
  • When an AI company retires a model, your agent tries its replacement instead of failing until the next update. Pictures made on OpenAI move to its newer picture model before the old one shuts down on October 23, and DeepSeek's Economy and Balanced move to the model that replaced its retired one.
  • Ask for a total from one of your lists and Archie adds it up itself, rather than leaving the sum to the AI.
  • The add-on store opens on tiles, one for each part of your life or work: Work, Money, Home, Health, School, News and fun, Writing, and Sales and clients. Press a tile for its add-ons, and an add-on for what it does and the button that adds it.
  • Archie Mobile is on the App Store. The code for getting the app now opens the App Store on an iPhone and downloads the app on an Android phone, instead of a page about it.
  • Every personality has a face of its own, and a personality's ears no longer get cut off in a reply.
  • Fixed: an email or text your agent read could approve a change it had asked you about. Calendar moves already waited for a yes you typed. Now removing something from a list, deleting a to-do, spreadsheets, documents, your CRM, connected services and adding an add-on wait for one too.
  • Fixed: a second question in a row about your calendar or mail got the same stock answer. It goes to your agent like any other message now.
  • Fixed: the welcome questions could keep going after you asked to skip them or asked for something else.
  • Fixed: buttons could arrive under a failure notice. They come with a plain question now.
  • Fixed: a message with no words in it got an error back. Now it gets no reply.
  • Fixed: the store's tiles overlapped on a narrow window. They keep four columns at every size.

Archie 0.3.6

October 7, 2026

  • Your monthly spending limit now also pauses reading new mail and texts, not only routines. A routine at the limit stays quiet instead of posting the limit message each time it's due, and your agent tells you once, at the end of an answer, when the month reaches your limit. Reminders you set still go off.
  • A shaky connection to the AI company no longer fails your request halfway. Archie sends it once more on a fresh connection. If it still can't get through, the message now mentions a VPN.
  • Asking for the same reminder twice at the same time sets one reminder, not two.
  • On iMessage, your agent can have an Apple ID of its own, which you text like a contact, so nothing shows twice. In the conversation with yourself, its replies now start with its name. The Connecting iMessage guide shows both ways.
  • Agents on Telegram, Slack, and Matrix can write tables again.
  • Telegram and Slack no longer preview a link your agent wrote, so neither opens an address before Archie has asked you about it.
  • An email your agent read for you can no longer shape a search in your next message. For about ten exchanges after it reads new mail or catches up on a group chat, a search it would hand on, or a site you name in passing, waits for your yes.
  • A note your agent saves after reading something you didn't write is labeled, so you can tell where it came from.
  • When your agent buys something, the total is read off the checkout page itself, and it won't press the button if the total went up after you tapped.
  • If Archie loses its place in a mailbox, it tells you which mail went unread, that every email is still in your inbox, and to ask your agent for it.
  • With your own AI key, a paused free trial no longer stops your 20 free jobs a day.
  • When a plan ends, the screen that asks for one has a Save a backup button.
  • A calendar add-on's Connect button asks Google for your calendar only, not your Gmail, and connecting an account again through a setup button keeps what it already had.
  • On the talk screen, click anywhere or press the space bar to cut your agent off.
  • On Windows, opening Archie again brings the open copy forward instead of starting a second one.
  • The store and the guides say the same things in about half the words.
  • Fixed: connecting Google failed on an agent that already had an Outlook account; the free jobs counter fell behind your replies, and one notice said "You was"; a one-line answer showed a fade under it; and long file names pushed the Knowledge tab past the edge of the window.
  • Fixed: deleting your account went ahead even when your plan couldn't be canceled. It now stops and says so, so nothing keeps charging.

Archie 0.3.5

October 5, 2026

  • Archie shows an alert on your computer when your agent posts something on its own, such as a reminder, a routine's report, or a card about new mail, while Archie is open behind your other windows. Turn it off in Settings, General. Archie Mobile does not buzz yet; a chat app such as Telegram is what buzzes your phone.
  • Type stop, or press Stop, while your agent is working. What it already finished stays done, and nothing more happens.
  • A send time your agent suggests for an email is a button on the card. Nothing is set to send until you press it.
  • Ask what your agent got done and it says what each job actually did, not only what you asked for, and whether each email it drafted went out.
  • Ask how many unread emails you have and your agent gives your inbox's own count. When a list it read was cut short, it says so.
  • Your agent says which calendar it reads: the main calendar of each account you connected, so it never calls a day clear. Before it changes your calendar, it names the calendar and the guests, and waits for your yes.
  • Your agent asks before opening a web address it came up with itself. It shows you the address first and waits for your yes.
  • Ask your agent how it works and it answers correctly: what it reads for you goes to the AI company along with what you type, Archie Mobile keeps a locked copy of your chat on our servers that only your phone can open, and it makes no compliance claim. When it does not know, it says where to look.
  • With a model running on your computer, Archie says what still uses the internet (your mail, your calendar, web pages and Archie's own checks) and names the app running the model when it stops answering. Moving your agent from that model to an AI company waits for your yes.
  • With Computer control off, ask your agent to buy or cancel something on a website and it names the switch that would let it, and says a purchase would still wait for your tap.
  • On a Mac with Liquid Glass, the sidebar and menus are glass and follow how see-through you set it in System Settings, under Appearance. Settings, General has a button that opens it.
  • A skill or a routine opens where it sits on the agent page: what it needs before it works, your answers, when it last ran and runs next, and Run now.
  • Getting an AI key goes one step at a time, ending on the box you paste it into. Every Google connect button tells you about Google's unverified-app screen before you see it.
  • On the talk screen, replies read aloud are kept short, it says what the microphone is doing, and a long answer scrolls. Turning on Let it talk no longer restarts your agent.
  • Fixed: the calendar showed the week the clocks change as half empty.

Archie 0.3.4

October 2, 2026

  • On a Mac, your agent can also use the apps you allow, not only websites. It presses buttons and fills in boxes inside an app's window without moving your pointer, so you can keep working, and if you start using that app, it waits for you. Ask for it in the chat: the first time it needs an app, it asks you, and macOS asks you once to allow it. It never uses Terminal, password keepers, System Settings, Mail, Messages, Calendar, or your web browsers, it never types a password or a card you pay with, and anything it could not undo waits for you. On Windows, it uses websites.
  • Computer control is one list. On the Connections tab, under Computer control, the websites and apps your agent can use sit in one list, with one Add box that takes a web address or an app's name.
  • You can see where your agent is working on a website. A small Archie pointer moves to each button it presses in its browser window, and rests on the one it is asking you about, so the picture in your chat shows which button it means. Your own pointer is never touched.
  • Fixed two cases where your agent could do more on a website than you had allowed.
  • Square Keeper reads your live Square: what came in, who hasn't paid an invoice, and when your money lands. It drafts and sends invoices and refunds a payment, each shown to you before it goes.
  • Social Posting posts to your Instagram, and to a Facebook Page if you have one. Send a photo and say what to say; it shows you the post and posts it when you say yes, now or at a time you pick.
  • Adding your own AI account takes less. Archie tells which kind of account it is from what you paste, and every agent you made on the free credits starts using it right away.
  • A stopped agent's chat says so, with a Start button, the way your phone shows it.
  • The Microsoft sign-in button says work or school account, and the sign-in buttons look like Google's.
  • Your agent types a gift card, library card, loyalty, or membership number when a site asks for one. A card you pay with is still always yours to type.
  • A text with a PIN in it is kept away from the AI, the way a sign-in code already was.

Archie 0.3.3

September 30, 2026

  • Fixed: an Outlook or Microsoft 365 inbox could stop being read and never start again. Your agent then saw none of the email that came in and could not draft a single reply. After this update it reads new email as it arrives, with nothing for you to reconnect. Email that came in while it was stuck is still in your inbox, but your agent does not go back and read it.
  • When Archie cannot read your email or your texts, your agent tells you in your chat. If it has not been able to read them for more than half an hour, you get one message saying so and what to do, or that there is nothing for you to do, and another when it works again. It only tells you to reconnect an account when the sign-in is really what stopped working.
  • Your to-do list is the first card on the Dashboard, unless something is waiting for you. It shows every open task. Tick one off (there is an undo), or type a new one straight into it.
  • Records are called Lists now, and Knowledge opens like folders. Documents, Lists, Memory and How you write each say under their name whether your agent only reads them or keeps them up to date. A document sits in the folder for its kind.
  • Your agent never presses a button that sends money or signs something on a website, even when you tap yes. That last press is always yours.
  • Reset also disconnects every phone and empties anything that was waiting to reach one.

Archie 0.3.2

September 30, 2026

  • Changing an email draft changes that draft. While a draft is waiting, ask for a new subject or a different line and the same card updates. You no longer get a second card with the whole email written again, and your agent knows the draft is still waiting for you. Nothing is sent until you press Send.
  • A reply to a draft card is a change to make. Reply to the card with "also say I might be ten minutes late" and the draft is rewritten to say it. To put in your own words exactly as you type them, press Edit it, then Use my words.
  • Rewrite it for me works on a new email, not only on a reply. Other versions is offered on replies, where it works.
  • Stopping something your agent watches happens straight away. "Stop watching that" or "never mind, I don't need it" marks it as stopped without asking whether you are sure. You can ask your agent to undo it.
  • Taking something off one of your lists asks first. Changes like checking something off or updating a price still happen right away, and your agent can undo them. A removal waits for your yes, because it is the one change that is hard to put back. A scheduled routine can change a list but never take anything off it.
  • Your agent checks your lists before saying something is not on them. When a message names something you are tracking, your agent is shown it, so it no longer answers that it has no record of it.
  • While the message box is empty, the microphone and talk buttons sit where Send goes.
  • Rules you wrote for your agent's personality stay with you when you remove the personality before choosing a new one. The next one you pick starts with them.

Archie 0.3.1

September 29, 2026

  • Your email drafts keep coming on a busy morning. When two emails arrived close together, Archie acted as if it had been switched off: it sent a "While I wasn't running" list and stopped writing drafts. Now each email gets its own draft, and that list only comes after Archie has really been off for an hour or more. Texts work the same way after a quick restart.
  • Fixed: old due dates coming back after the computer was off. A reminder that was due while the computer was shut, or had lost power, ran late and forgot what it had already told you, so it repeated past due dates as if they were new. Now it only mentions what has newly come due.
  • Reminders say what is due, and the month. "Your Honda's registration expires on October 19", not "Registration expires on the 19th". A savings goal whose date has gone by says so, and asks whether to set a new date or close it.
  • The morning brief shows when each overdue to-do was due, and tells you to say which ones are already done so your agent can check them off.
  • The last button on your task list says Got it. It used to say "Leave the list", which looked like it would delete your tasks. It never did. Pressing it puts that message away, and anything unfinished shows up again on the next list.
  • A power cut is much less likely to cost your agent its memory. Archie now finishes writing each file to the disk before it replaces the old one, so losing power partway through a save no longer leaves an empty file behind.
  • Your agent sees more of what you said. Reports from routines used to fill the recent conversation your agent reads, and push your own messages out of it. Only the newest two reports stay in it now, and anything pushed out is still checked for things worth remembering. Forget this also reaches older memory, so a fact you took out stops coming back.
  • A reply that gets cut off says so. If your AI company stops answering partway, or the internet connection drops, you get the words that already arrived and a line saying what was done and what to check.
  • Copy chat copies the whole conversation as text, and Cmd+A (Ctrl+A on Windows) selects just the chat.
  • On a Mac, Archie looks like the rest of macOS 26: the system font, a sidebar you can see through, and the same colors as the iPhone app. The red, yellow and green window buttons come back in full screen.
  • The buttons under a message sit together in one bar, each in its own color, so Send later no longer looks like a second Send.
  • The agent page is shorter at the top, every skill and routine row has its own on switch, and jobs say Running, Done, Could not finish or Waiting for you.
  • Settings and Account are down to nine sections, with icons. You pick a theme from pictures, Spending shows bars, and pairing a phone is on an agent's Setup tab, under Chat app, on the Archie Mobile row.
  • Archie Mobile can be paired from an Android phone.
  • Signing in says what is happening at each step, and the page your browser lands on afterward matches light or dark mode.
  • Plainer words on every screen, in error messages, guides and the store, and in what your agent says. Three routines have new names: Daily Briefing is now Today's Calendar, Meeting Sync is Meeting Recap, and Meeting Desk Sync is Call Filing.
  • New agents start with a shorter, plainer personality. Agents you already have keep theirs.
  • New in the store: the Money at Home pack. This is Archie; Archie for Business does not carry it.
  • Archie opens older Excel files (.xls) and OpenDocument spreadsheets (.ods).
  • Catch-up lists show the real time each message arrived, say which inbox when you have two, and put whole names on the buttons.
  • On Telegram, a question written under a file you send gets answered, and a message that waited for the computer to wake up is answered knowing it waited.
  • The inbox button on the first run now sets up Email Manager, so the morning check it offers actually runs.

Archie 0.3.0

September 22, 2026

  • The answer turns up as it is written, a line at a time, instead of arriving all at once when your agent has finished thinking. This is in the Archie window, on an Anthropic account of your own. On the other AI companies, and in Slack, Telegram and the rest, the answer still comes in one piece.
  • Your agent writes a real PDF. Typeset on this computer, with numbered pages, and if the answer had a chart in it the chart is drawn into the page rather than listed. Ask for slides and you get a deck with one heading per slide. The deck opens in Keynote. Nobody here has been able to open it in PowerPoint yet, so it does not promise that until somebody has.
  • It can read a PDF that is a photograph of a page. Some PDFs are pictures of paper, with no words in the file at all, and Archie used to say so and stop. Tick Read PDFs that are scans, under Setup, and it looks at the pages instead: up to ten of them, when you ask it to open the file and there was nothing to read. It costs a cent or two each time, on your own AI account, and it needs an Anthropic key. With any other AI company the box does nothing and the screen says so.
  • Paste in an AI key and Archie opens, free. No plan, no trial running out. You get 20 jobs a day, five of them held back for things you ask for yourself, so a routine cannot spend the lot before you wake up. The count starts again at midnight. This is Archie; Archie for Business still takes a plan.
  • You can switch off things your agent can do. Under Setup, What it can do, there is one switch for each: building skills, adding and removing add-ons, running routines, setting reminders, reading its own record, saving files, changing its own AI settings, and going to the web. All of them start on. Turning one off stops your agent offering it, and the row says where to turn it back on.
  • Your agent keeps what it used to forget, and you can tell it a fact yourself. What it remembered was trimmed oldest first, so something you mentioned once in March was gone by September with nowhere to look it up. What gets dropped now goes to an archive beside it and comes back on its own when something you say touches it. The Memory panel takes typing now too: write a fact in, or correct one it got wrong.
  • Two tabs have better names. Now is called Dashboard and Work is called Jobs. The messages your agent sends you say the new names too.
  • Connecting Gmail asks Google for one permission instead of three. It was always one permission; the consent screen just described it three times, which read as three.
  • Replies start about a second sooner. Every message waited 1.2 seconds before any work began, holding the door open for a second message that in most cases was never coming. The wait is a quarter of a second now, and the full wait comes straight back the moment a second message proves you are still typing.
  • Every service in the store carries its own mark, and you can narrow the store to one app you have connected and see only what works with it. Five marks came off because the companies that own them say we may not draw them; those rows say the company's name instead.
  • Archie for Business has its own icon, in the dock, in the window and on the download.
  • Archie has a trademarks screen, naming the companies whose marks appear in the app and saying the marks belong to them.
  • New in the store: My Classes, Class Planner, Leave Time and Flight Check-In, with routines to go with them: The Week in Classes, Due Soon, The School's Own Dates, The Morning Drive Check, The Traffic Watch and Check-In Window. There is a College pack that puts the class ones together. The old Student pack is now called Self-Study, which is what it always was; nothing in it changed.
  • Ember has a back now, four more things it does between answers, and a backflip.
  • The box you type in grows with what you are writing instead of scrolling inside four lines.
  • Paste a new key straight over a connected app, instead of disconnecting it first and starting again.
  • A routine you have switched off now looks switched off. The only sign used to be a small pill at the end of its name, with the line underneath still reading "weekdays at 7:00 am".
  • Fixed: ask about one mailbox and your agent reads that mailbox. The list it worked from was capped at ten messages across every address you have connected, and the cut was by how recent they were, so one busy address could take the whole list and the address you asked about contribute nothing to it. Each one is guaranteed a share now, and your agent can tell you there is more to see.
  • Fixed: the browser window your agent opens now closes whichever way the job ends, and Archie comes back to the front when it hands something over to you.

Older releases

Archie 0.2.5 back to 0.1.1, as their notes read when each one came out. Some names and prices in them have changed since. Press a release to open it, and see Pricing for today’s prices.

Archie 0.2.5September 15, 2026
  • Your agent is on your phone. Pair one from Settings and the same conversation is there: type to it, speak to it, hand it a file, and hear it answer out loud. What you type at your desk turns up on the phone too, and a message sent from the phone says it came from the phone.
  • The phone runs your agent, not just talks to it. Rename it, change its face, switch a skill off, build a new one by hand, change how much thinking each reply gets, and read the Work board. Anything it has drafted and is waiting on you for shows up there to read.
  • A Now tab, above Chat, and the agent opens on it. One screen for what needs you today: drafts waiting on a press, replies you are owed, reminders coming up, and what your agent is doing this minute. A draft whose moment has passed drops off it instead of sitting there.
  • A Calendar. Your week in time order, putting together the calendar you connected, the reminders you set and the routines that are due, with a mark saying which is which. Until now the connected calendar was never drawn anywhere in the app.
  • Your to-do list has a place of its own. It was in the Records table alongside people and orders, three presses in and with nothing on it saying what it was. It is a tab now, open tasks first, in the order they are due.
  • Ask your agent to switch AI and it switches. "Use ChatGPT instead", "think harder from now on", "keep your answers cheap". If there is no key saved for the company you asked for, it tells you where to paste one instead of just saying no.
  • Running out of AI credit now says what to do about it. The message names your provider's own billing page, the Start button says the same thing instead of printing the provider's error back at you, and the AI account card keeps that link where the key lives.
  • Rooms: several of your agents in one conversation. Name one and it answers, the rest stay quiet. New room is in the agent list, under your agents.
  • Scheduled runs and the inbox watch cost about half what they did. Work nobody is sitting waiting on goes out on the cheaper lane, and a second message in the same chat reads what the AI company already had rather than paying to send it all again.
  • Outlook works the way Gmail does. A first look reads the mailbox and tells you what is in it, and your agent can file a mail into a folder you named on any of the three providers.
  • Getting started is one screen with things to press on it. Three asks to choose from instead of five counts to read, the name and the face on one card, and your agent arrives wearing Ember's face and named for it. The walk around the page waits until the interview is over rather than talking across it.
  • A skill that arrives needing something walks you through it instead of sitting there doing nothing, and Build a skill is a place you go under Skills, with a conversation of its own.
  • Your agent shows everything a turn did, not only its last step. A turn that is taking a while says what it is doing at 45 seconds, 2, 4 and 6 minutes, and how to call it off. It stops on its own at eight.
  • The draft you had before an edit is one press away, and saying "edit" opens the same editor that pressing Edit does.
  • New in the store: Lead Desk, Writing Desk, Ready to Post, Common Questions, a Stripe digest, Job Hunt and Application Sweep.
  • Ask for the same reminder a third time and your agent offers to make it a standing one rather than setting another copy.
  • The app says Beta on the screens that carry the wordmark, because that is what it is.
  • Fixed: a numbered list from your agent came out as 1, 1, 1.
  • Fixed: the card asking one thing of you took 25 seconds to turn up after your first answer.
  • Fixed: an icon and its word on a button could end up on separate lines.
Archie 0.2.4September 8, 2026
  • Your agent can read your Zoom calls, with no Zoom account and nothing to connect. When you record a Zoom meeting, Zoom saves it in a folder on this computer. Archie reads what is in that folder and writes up what was said, here, with nothing sent anywhere. Turn it on under Connections, then Apps, then Zoom. Only calls you host are saved that way, and only the ones you pressed Record on, unless you have turned on "Record automatically as I start a meeting" in Zoom.
  • Zoom often writes the transcript itself, and when it has, that is what Archie reads: it says who said which line, it takes no time, and it uses none of this computer's power. A call with no transcript beside it is transcribed here instead, and then the write-up says what was discussed and who was on the call but never who said what, because the audio does not carry that.
  • Meeting Notes reads whichever meeting source you have, and any one of them is enough. A recording on this computer, a Fireflies key, or both. The card that asks for one no longer reads as though Fireflies were the only answer.
  • Your agent tells you what it is costing you. A week after you start it says what it has spent so far and what that works out to over a month, then once a month after that. It says nothing at all rather than guess when it cannot put a price on something.
  • Ask about your week and it reads every calendar you have connected, not just the first one. Booking something asks which calendar it should go on, when there is more than one to choose from, and the confirmation says which one before you press it.
  • A list can point at another list. When you build a list on the Records tab, a text column can name the list it points at, and then you pick a row from a menu instead of typing its name. Tasks now point at Projects that way, so "what is left on the kitchen" is a question with an answer.
  • A routine's report can ride the next one. Switch on "Fold into the next routine message" and that routine goes quiet and arrives underneath the next one that speaks, so a morning is one message rather than four. Nothing is lost if there is no next one: it goes out on its own after six hours.
  • Text Replies can watch group chats. It is off until you switch it on, under Which texts to watch, because a reply you send from a card there goes to everyone in the group. With it on, a card says who spoke and in which group, and a promise you make in a group is remembered against the person you made it to.
  • A promise you make in a text is now on your Commitments list, not only a reminder. So it can be asked about, chased on the days the sweep keeps, and closed, the same as one you made in an email.
  • "Who owes me a reply" counts what you asked a group, not only what you asked one person.
  • The Connections tab's Apps list now says which half you are looking at: what this agent already reaches, then a line, then what it could. The first tile of the catalog used to read as one more connected app.
  • Zoom, Cloudflare and the MCP servers have their own marks on that list instead of two gray letters.
  • On the plan with the AI included, your agent will not read your texts, for the same reason it will not read your mail: on that plan it thinks on our account, and the person who texted you never agreed to that. Connect an AI account of your own and both start working, with nothing to switch. The note with 0.2.3 said email was the only thing that plan does not do. It is both.
  • Archie writes dates the way your agent does now, month first, everywhere. Before, the app asked this computer how to write a date and the agent in the next pane wrote it the other way round.
  • Getting started is shorter to read. The Setup tab shows what is needed to run and folds the rest to a line, the AI account pane offers three companies and a door instead of a wall, the box you type in has one button for adding a file or a picture, and the first screen leads with one line instead of five counts.
  • When you let your agent interview you, what you say now fills in its settings, and it tells you what it filled in.
  • Websites shows up on the Connections tab once something actually asks for it, instead of sitting there from the first day.
  • Fixed: a reply to a mail that copied other people went only to the sender. Answering an introduction reached one of the two people in it, and only a note in brackets said so.
  • Fixed: a second message in the same chat costs a fraction of what it used to. The line telling your agent the time was moving on every request and throwing away everything the provider had remembered; a chat's second turn now bills at about a fifth of what it did.
  • Signing in records that you agreed to the Terms, with the version and the moment, and says so under the buttons.
Archie 0.2.3September 3, 2026
  • Archie now comes with the AI included, if you want it that way. $59 a month or $599 a year buys the same plan plus $25 of Claude usage every month, so there is no AI account to open and no second bill. Tick "With the AI included" when you pick a plan. Your agent runs on our Anthropic account, which means what you write to it passes through our server on the way to Anthropic; nothing there writes it down, we keep no copy, and we use none of it. Email is the one thing this plan does not do: your agent will not read or write mail while it is thinking on our account, because the people who write to you never agreed to that. Connect an AI account of your own and every email feature starts working, with nothing to switch. When the $25 is used, your agent pauses until next month.
  • On that plan, and on the free credits, the Response quality control on an agent's page now says that it changes nothing until you connect an AI account of your own, instead of looking like a setting that does something.
  • Any service that publishes an MCP server can be connected now. GitHub, Linear, Stripe and Cloudflare are on the Connections tab under MCP servers, and "Another MCP server" takes any address you have. Paste the key, Archie asks the server what it can do, and your agent can use every tool the server offers: the read-only ones on the spot, and anything that changes something only after you approve it, the same way a web request works today.
  • Ten more services have an add-on that knows how to use them: Airtable, Calendly, GitLab, Lunch Money, Mercury, Readwise, Stripe, Splitwise, YNAB and Zotero. Each one was written from that service's own documentation, so your agent is told the real endpoints instead of guessing.
  • A key you paste for a service no add-on uses works anyway. Before, "what is on my Calendly" was answered with "Calendly is not connected" while the key sat right there.
  • A connection assigned to somebody by name on a shared agent is now visible to them. Before, a teammate who asked whether their own Calendly was connected was told no, though it was.
  • A routine that missed its time because the laptop was shut says so. The Routines card names the sleep and opens a list of steps for keeping this computer awake, instead of sending you to look at Archie, where nothing is wrong.
  • A morning routine that missed 8:00 am now runs any time before noon, an afternoon one runs until 6:00 pm, and an evening one runs until midnight and never after. The old rule gave up two hours after the time, which on a laptop opened at 10:00 am meant last night's routines never ran and this morning's barely did.
  • The Routines tab checks how each routine is doing every few seconds. It used to look once when the page opened, so a routine that had recovered could stay marked red all day.
  • Remind me to pay Sam back on Friday sets a reminder for Friday at 9:00 am and says so, with Noon and 6:00 pm one tap away, instead of stopping to ask for a time. "Pay my electric bill on the 5th" sets it for the next 5th and asks whether you want it every month; a Yes turns it into a routine, and a repeating reminder is a routine now, with no skill needed.
  • Tell your agent what you wish it could do and it asks one question at a time, as buttons, about your life and never about the machinery. The card says what it assumed. It refuses to invent a time, a time zone or a name you did not give it.
  • "Every Friday morning, tell me what bills are due" sets up a routine over a skill you already have, offering a store routine first when one fits. Run it now is one tap after Keep it.
  • A store routine runs the skills it came with, so its edit form says what it runs instead of offering tickboxes. Only a routine you built shows the checklist.
  • When your agent has a document in its knowledge base, it reads it and answers, instead of telling somebody on Slack to open a file they cannot see.
  • On the edition that answers one person, the two buttons for adding a second are gone. Both could only produce an error.
  • Fixed: a network failure said "network request failed" and nothing else. It now says why: the name did not resolve, the connection timed out, or the secure handshake failed, which are three different things to go and fix.
  • Fixed: a document over the size limit was cut twice, losing the end of it a second time.
  • Fixed: the crash reporter could turn a recoverable crash into Archie quitting outright when the terminal it was writing to had gone away.
Archie 0.2.2September 2, 2026
  • Your agent can now read the mail at an iCloud, Fastmail, Yahoo, AOL, Zoho Mail or GMX address, and write replies for you to send. Pick the provider on the Connections tab, make an app password on that provider's own website (Archie links you straight to the page), and paste it once. iCloud, Fastmail, Yahoo and Zoho bring their calendar with them. Everything already true of Gmail is true here: a reply goes out only when you send it or set a time, and mail it files is moved to Archive rather than deleted.
  • Newsletters and mailing lists stop being read by the AI. Archie used to skip them using Gmail's own Promotions and Social labels, which no other provider has, so an Outlook or iCloud mailbox was paying to have every newsletter read. It now recognizes bulk mail from the message itself, on every provider. A receipt from a shop is not bulk and still reaches you.
  • Tell your agent what you wish it could do, and it writes the skill. Say it in your own words in chat, and it comes back with a card: the name, what would set it off, and which of your accounts it would use. Nothing is added until you press Keep it.
  • Every add-on is included with Archie. The Buy button, the prices on the cards, and the block that used to sit in front of Meeting Notes are all gone, because there was never anything to pay for.
  • Archie can keep a backup for you, once a week, on its own. Choose a folder once and it writes one every week and keeps the last three. It writes the first one the moment you switch it on, so a folder it cannot write to is a problem you find now instead of in seven days.
  • A backup now knows which computer wrote it. Restore it on the same machine after a mishap and nothing is disturbed. Restore it on a new machine and Archie clears the things that cannot survive the move (the browser sign-ins, the Signal link, the lights on the old network) and lists each one for you to reconnect, instead of leaving them looking fine and dead.
  • Saving a backup asks whether to stop the agents on this computer. That is the moment somebody moving to a new machine is standing in front of the old one, and two computers answering the same mailbox is the thing nobody notices until both have replied.
  • Signing in on a new computer asks about your backup first. Before, it built you a new agent and the restore was behind a box asking you to type "replace" over the agent you had just been given. If you are on the free trial you are not asked, because there is nothing to restore.
  • Every row in the reconnect list is a button now. Each one said where to go, and none of them went there.
  • Editing a draft asks what you meant, because one button could not mean both things. Rewrite it for me takes what you type as an instruction. Use my words puts your text in as the reply, exactly as you typed it. Before, "I also had one at 12 and am still on call with Jack" could become the whole message to your friend.
  • A drafted reply capitalizes the way you do. Archie had been told to write in lowercase where it looked natural, and that instruction was beating the samples of your own writing sitting right next to it.
  • A reply knows who it is from and who is copied in. An email saying "I am connecting you with Jett" produced a draft from Jett offering to connect with Jett. Cc is now carried through, kept separate from To, because being copied in is a different fact from being written to.
  • Pointing Archie at a model running on another computer works properly. The Connect button no longer demands a key that Ollama does not have, the check that catches a model which talks but never uses a tool now runs on a typed address, and an address that will not answer names the three things that are actually wrong instead of telling you to check your internet.
  • Fixed: saving your own provider and pressing Use it handed Archie the model name where the key belongs, and left the address blank. A second agent in the same workspace then refused to start.
  • A skill you write yourself can ask for your mail. It could already reach your calendar and your tasks, and nothing on the screen said why the inbox was missing.
  • A Home Assistant address that cannot work now names every reason it cannot, including the one for people reaching their house over Tailscale.
  • Fixed: two things finishing at the same moment could shred the record of what they cost, so the spending page lost both lines.
  • Fixed: one long column pulled the other three headings off the side of the window.
Archie 0.2.1August 30, 2026
  • Your agent can now do a job on a website end to end: sign in where you sign in, click through, download what it came for, hand it to you, and keep it. Most of this release is that one job, found by collecting a real bank statement over and over until there was nothing left to go wrong.
  • The browser window opens on your Mac's own chip. It was launching the Intel build through Rosetta, which is the whole reason the window it opened felt slow.
  • The Websites panel says whether each site is signed in, and when it last was. It said nothing, so a site that had signed you out looked exactly like one that had not.
  • Closing the sign-in window is noticed. The panel used to keep saying the window was open, because it watched for the program rather than for a page, and closing the last window does not close Chrome.
  • "A window just opened" is now said when a window has opened, rather than when the program was still starting.
  • Your agent stops reloading over your sign-in. Asked to open a site it was already looking at, it went back to the front page, which on a bank signs you straight back out.
  • A file a site hands over lands where the job can find it, and the job waits for it to arrive. It was being told where to put downloads one tab at a time, so four or five statements went to a folder nobody looked in, and a file still arriving when the browser closed was reported as one that never started.
  • The same file handed over twice arrives once.
  • "I'll do this bit" ends with a button. You had to type "done" at an agent that had just given you buttons for everything else.
  • Your agent can read back the question it asked you. It sends a question as the caption on a picture of the screen, and could not see that caption afterwards, so your answer meant nothing to it and the job went round in circles.
  • It stops handing you the clicking. Asked to sign in, it now names the window already open in front of you, instead of a screen inside Archie that does not exist.
  • A browser job gets the room to finish. Every job stopped after ten steps, and for a browser a step is a click: one bank statement spent all ten opening the site, clicking through, downloading the PDF, sending it to its owner and filing the record, and then said "Reached the tool call limit" underneath the statement it had just collected.
  • A job that does run out of steps now says what it finished and what is left, rather than reporting the whole thing as lost.
  • The "let this agent use websites" warning has a button that takes you to the switch.
  • Pictures and files arrive in the Archie window, not only on your phone. A statement your agent had just written to a folder six inches away showed up in the window as a sentence telling you it had gone to Telegram.
  • Your agent can read a file it downloaded for you. It fetched a bank statement, sent it over, and twenty minutes later said it could not see the Downloads folder and asked for the file to be uploaded.
  • A PDF whose text cannot be read says so instead of guessing at it. Some PDFs come apart into nonsense that still looks like words: in a real bank statement every date and amount arrived scrambled, and the agent decoded most of them and got four wrong, including ten dollars going out that became a hundred and ten dollars coming in. Those lines are now taken out and named as missing.
  • A total in a spreadsheet your agent writes is the sum of the rows it wrote, never a figure copied in from somewhere else.
  • Your agent can change one number in a spreadsheet you already have. "Add this to my expenses" used to end in a second spreadsheet and an evening merging the two.
  • A spreadsheet it reads says what it cannot see. Colors, merged banners and frozen rows are how a real sheet tells three tables apart, and it was answering as though the numbers were the whole truth.
  • Statement Collector collects each account on the day that account's statement is ready, instead of doing the whole round on one date. A card closing on the 3rd and one closing on the 27th are three weeks apart, and neither is late. It works the dates out from the mail that already arrives, so there is no list to keep.
  • It asks once what you want done with a statement, with buttons: a running spreadsheet, a summary of what changed, or both. Then it remembers and never asks again.
  • Asking your agent to forget something now forgets it. It had a button for this on one screen and was told in as many words that it had no way to do it, so "forget that I play Rocket League" got directions to a screen instead. Anything it forgets can be put back.
  • Your agent picks things up as you talk, without being asked to remember them, and a hobby it picked up can no longer push a deadline out of its memory.
  • Something a correspondent said about their own life can no longer end up in your agent's memory of you.
  • Two texts sent one after the other get one reply covering both. The card answering the first was retired half a minute later by a card that answered only the second, so the first answer was not wrong, it was gone.
  • A reply drafted for you sounds like you, not like the person who wrote to you.
  • An account on the Connections tab says whether the mailbox actually opens, not only what it was allowed to do. A withdrawn permission, a changed password and a revoked account all looked like eight green ticks on a mailbox that had not opened in a week.
  • You can say which model each of the three response quality settings uses, if you are paying per token on your own key and already know which one you want.
  • The spending page stopped telling people on DeepSeek and Mistral that they were spending nothing.
  • Groq is described as the cheapest option rather than the free one. Its free tier cannot run a single Archie turn, and we were recommending it first.
  • Fixed: a GPT-5 model left thinking spent its whole reply budget on thinking and answered nothing.
  • Fixed: typing /new in the Archie window looked like it hung. The confirmation had already arrived and was sitting one line above the message asking for it.
  • Fixed: a picture or a note sent while your agent was still writing appeared under the reply instead of above it, so you read the answer, then the button you had already tapped, then the question that button was answering.
  • Asking your agent whether the store has something now searches the store. It answered from memory, or said it could not look.
  • An add-on installed from chat says when it is ready. It said "ready in a few seconds" and then went quiet, because what makes it work is a restart.
  • Fixed: saving as you type was restarting your agent as you type.
  • Fixed: the Back arrow could not undo moving to a section of a page, only to another page.
  • Fixed: an error screen threw away the half of the message that said what happened, and stayed after the problem was gone.
  • Fixed: one long line in a routine could push a whole tab off the side of the window.
Archie 0.2.0August 29, 2026
  • Sixteen people who had never seen Archie were written down as people, not as a feature list, and 219 things they would really type were pushed through the product one at a time. Everything below is what that found. Eighty eight problems, all of them closed, and then a second pass by readers who had not been told what was fixed.
  • Your agent knows it is shared. On an agent with more than one person on it, it was never told that, so "can Sam see this?" and "who else reads these?" were answered from a guess. It now knows who the owner is, who is speaking, and that the owner can read the conversation, and it says so when somebody asks.
  • Everything your agent sends without being asked now goes to a chat you picked. Routine reports, reminders, watcher cards and morning briefings all went to whichever conversation happened to speak to it first, which on a shared agent could be a colleague's.
  • Approving somebody from the waiting list now sends them a message. They were let in silently, so the first thing a new person heard from your agent was the answer to something they had asked before they were let in, and nobody ever told them the owner can read what they write.
  • You can ask your agent whether a routine actually ran. "Did my morning brief go out?" and "has the pipeline report ever worked?" now have a real answer, including "it has never run", which is usually a computer that was asleep at the scheduled minute rather than anything broken.
  • Switching a routine off by asking now switches it off. It reported success and kept firing until Archie was restarted.
  • Removing a skill now takes its routines off the clock with it. They kept their schedule, fired every morning doing nothing, and stamped themselves as having run.
  • A routine you add by asking for it keeps your time zone. It was set in UTC, so "every weekday at 7" meant 2am for somebody in Chicago, and the confirmation never named the hour.
  • A routine whose time passed while the computer was asleep is delivered late instead of thrown away, and a routine that is broken says so on the second failure rather than after two more weeks of its schedule.
  • You can talk to your agent in the Archie window and be answered out loud. A voice note recorded there came back as a paragraph to read, while the same voice note sent from a chat app was answered in speech.
  • A spoken "go ahead" now lands on the card you aimed it at. Recording a reply to a proposal reached your agent aimed at nothing, so the change you had just approved was already gone.
  • A long spoken answer says why it came back in writing instead of going quiet, and a reply read aloud keeps the options in it. The one thing read out was the version with the choices stripped out, so somebody driving heard a question with no answers.
  • "Read that back", "say that out loud" and every other way people ask for speech now work. It matched fourteen exact phrases and nothing else.
  • You can say "send it" instead of tapping. On the chat apps that draw buttons, the words were switched off, which is exactly where somebody with their hands full needs them.
  • A draft your agent scheduled for later cannot go out on its own without you asking for that time. The prompt promised nothing leaves without a press, and one setting had been quietly sending mail. A scheduled send now needs a person on the other end of the conversation and a time at least ten minutes out.
  • A reply that failed to send is no longer left in the conversation as something your agent said it sent, and a failed delivery no longer destroys the reminder or the card it was delivering. Both go back on the list.
  • Your records save out as a spreadsheet file. On the agent's page, under Records, there is an Export button that writes a CSV you can open in Excel or Sheets. The only way out before was asking your agent to retype the rows into the chat, which is slow, costs money by the row, and cuts off partway with nothing saying it did.
  • Archie can tell you what removing it leaves behind, and where. Nothing in the product could say that your profile, memory, transcripts and job history stay on the disk after the app is gone.
  • Your agent can say its own name. The name you gave it never reached the part that writes its replies.
  • Ask "what did you do this week" and the answer comes from the record rather than from memory of the conversation. Two records existed and disagreed with each other.
  • Every job on the Work board says who asked for it. Work a teammate requested was labeled "You asked" on the owner's board, and the name had been recorded for exactly that purpose and thrown away.
  • A job that asked you a question in words stops sitting under "Needs you" forever. There was no card to press, so nothing could ever clear it.
  • The store search finds add-ons by what you would call them. It could not see the trigger phrases every add-on is written to be found by, and answered a miss by saying the catalog cannot do it.
  • An add-on your agent installs from chat works straight away. It was inert until Archie restarted, and the reply said it was ready.
  • Add-ons that cannot install are no longer offered. The offers reached the model with every requirement stripped off them.
  • A skill you write yourself with a blank "One line" is no longer invisible. Every screen said it was installed and working while nothing could route to it.
  • The skill editor names the tools you can use, and says which field decides when your skill runs. It asked you to name tools and listed none.
  • A reminder set by voice arrives out loud. It was delivered silently while the description your agent wrote its promise from said it would speak.
  • A reminder Archie set for you off the back of a text says so, and names the switch that turns it off. It appeared in your list with nothing to say where it came from.
  • Times are written the way you would say them, on a 12-hour clock, including the read-back whose whole job is to let you catch a wrong hour.
  • Your agent's memory is honest about what it holds. A background pass was writing durable notes about you with no tool call, no record and no sentence in any reply, and one standing instruction told it not to say it did not know.
  • Undo covers the whole turn. Several deletes in one turn shared one undo slot, so it restored one row and your agent confirmed everything was back.
  • Fixed: a turn that died left "Working on it" in the thread forever. It now owns up to it on the next start.
  • Fixed: a delivered text could report itself as failed and invite you to send a second copy.
  • Fixed: undoing a filed email did nothing on Outlook. It worked on Gmail only, and said nothing.
  • Fixed: a guest could tick the owner's task off, or settle the owner's card, by typing a number instead of tapping.
  • Fixed: a guest's words could be written into the owner's permanent memory.
  • Fixed: on a shared agent, a group room could be adopted as "the owner's own conversation", which put everything the product keeps out of rooms into one.
  • Fixed: every check that keeps a room out of the delivery chat was inert on Discord.
  • Fixed: /invite on an agent that only talks in the Archie window minted codes nobody could redeem, and told the owner to hand them out.
  • Fixed: clearing a conversation was open to anybody in it, so a guest could wipe the room's thread for everyone.
  • Fixed: connecting a company mailbox gave your agent no mail ability at all, and nothing said a skill was also needed.
  • Fixed: the read-only calendar note and two other connect sentences were written to the owner and read out to guests word for word.
  • Fixed: a table asked for in the Archie window rendered as literal pipe characters.
  • Fixed: a stopped agent drew no Start button once there was a conversation on screen.
  • Fixed: two evening journal routines fired half an hour apart on every new agent, one of them the add-on written to replace the other.
  • Fixed: the alarm for a hand-written skill naming a tool that does not exist was sent to us and never to the person who wrote the skill.
  • Fixed: a records query that matched nothing looked exactly like a list that is empty, and a query capped at fifty rows read as the whole list.
  • Fixed: a document read from your knowledge base reached the model with nothing marking it as something a stranger wrote.
  • Fixed: every turn told your agent its voice had just changed, on agents whose voice never had.
  • Fixed: the first-run interview swallowed every message you sent, including requests, and had no way out from a phone.
  • Fixed: your agent was told it could not do things that were on its own tool belt, and told to point at instructions that were not there.
Archie 0.1.9August 27, 2026
  • Your email signature belongs to an address now, not to the whole agent. On an agent your team shares, a reply written from a colleague's own address went out over the owner's name, and the one person who could see that happening was the client reading it. Every address is signed with its own signature. An address that has none is signed with the name of whoever it belongs to.
  • A signature can be a picture. Put a logo beside the words, or use an image that is the whole signature with the name, title and phone drawn into it. You pick small, medium or large, and the preview is drawn at the size that goes out rather than as a thumbnail, so you find out it is too big before a client does.
  • Somebody who is not the owner can set their own signature by telling the agent in chat. The Archie window belongs to the owner, so until now the owner was typing everybody's signature.
  • A draft can wait for a time you pick. Every draft your agent hands you has a Send later button offering an hour from now, three hours, or the next morning, each printed as the time it actually comes to. If the other person writes back on that thread before it goes, the queued reply stops itself and hands the draft back to you with a line at the top saying why. Nothing leaves your account unless you send it or set the time for it yourself.
  • Your agent can start an email, not only answer one. "Send Sam an email saying let's get lunch" used to come back as an explanation of why it could only reply to mail already in your inbox.
  • Campaign Watch is new. It reads your lemlist campaigns and tells you who wrote back, what they said, and which campaign has quietly stopped working. It writes the replies and you send them.
  • On a shared agent, somebody who is not the owner can answer mail sent to their own address. Every mail action counted as acting on the owner's accounts, so the only way to let a colleague reply to their own client was to hand them the owner's mailbox as well.
  • A calendar can now be shared while the mail on the same account stays private. Sharing somebody's working hours, which is what colleagues need from each other all day, used to mean sharing their inbox too.
  • Connected apps say whose they are: one person's, or the whole business's. There was one key per service, so "add that to my list" meant the owner's list wherever the person asking sat.
  • Meeting notes accounts are one per person as well. Pasting a second Fireflies key used to replace the first with nothing on screen to say so, and whoever pasted last was the only person whose meetings the agent could see.
  • The owner can read what the agent has said to everybody else on the roster, on the agent's page. Every conversation but the owner's own happened somewhere the owner could not see. It is read only, with no box to type into: answering in somebody else's chat would put the owner's words in a room the owner is not in.
  • Everything you reach with a pasted key is one list on the Connections tab, under Apps. Meeting recorders had a tab of their own and GoHighLevel had no place on the tab at all, so connecting a CRM meant going and finding the add-on that wanted one.
  • Reconnect opens inside the row you pressed it on. It used to open a panel at the top of the list, so pressing Reconnect on the fourth mailbox opened something over the first one, and the only way to tell what you had pressed was to read the address in the heading.
  • A connection that nothing installed can use now says so, with an amber dot on the Connections tab and on the sidebar beside it. The add-ons that would use it are buttons you can press right there, instead of a trip to the store to look up what you have just been told.
  • Fireflies, lemlist and GoHighLevel have their own logos in the list. They wore two-letter monograms beside twenty real logos, which read as three services that had failed to load.
  • Response quality now reaches the part of your agent that reads mail and texts in the background. That part was pinned to the middle setting whatever you had chosen, so somebody who moved the dial to Economy to spend less went on paying the middle price on every message that arrived, and no screen said so. Deciding what a message is amounts to sorting the mail, and sorting is what the cheaper settings are for.
  • There is a second number beside "Tell me about at most", called "Read at most". The first one is about your attention. The second is about your money, because working out what a message is costs a little whether or not it turns out to be worth telling you about, and nothing counted that before. Neither one ever holds up a Draft button you pressed yourself.
  • The Spending screen can name what used to land in "Everything else", which was about half of every call your agent made. A row called everything else is only useful when it is the small one.
  • Your agent pays less to read the same mail. A reply carries every earlier message in the thread underneath it, and your agent was being handed all of it every time. The quoted part comes off first now, so what it reads is the message somebody actually just sent.
  • Routines that come due in the same minute run one after another instead of all at once, which makes every one after the first cheaper.
  • The chat says what a model sounds like when it is wrong, under the box rather than in the transcript. Being right and being wrong sound identical coming out of one: no hedge, no hesitation, nothing in the words to go on.
  • Fixed: Response quality had quietly stopped existing for every add-on installed from the store. It survived only on skills you write yourself, so a setting the app describes in three places was adjustable on almost nothing.
  • Fixed: an Outlook mailbox could report itself as broken when it was only busy. Outlook allows one app four requests against a mailbox at a time, and a twenty-message list went out as twenty at once, so the mailbox refused and your agent handed you the refusal as an account problem and named the address, which invited you to reconnect something that was never broken.
  • Fixed: an add-on that names the one service it needs was handed every key on your agent instead of that one. An add-on that names nothing still reaches everything, which is what an add-on you write yourself does.
  • Fixed: starting a new agent from an existing one copied connections belonging to somebody who is not on the new agent, which left keys nobody there could use.
  • Fixed: the remove button could fall off the end of a mailbox row.
  • Fixed: on Windows, the video add-ons downloaded an old copy of one of the tools they need, so they did not run.
  • Fixed: on Windows, setting up phone access rebuilt everything every time it checked, because it could not tell that nothing had changed since the last check.
Archie 0.1.8August 26, 2026
  • Your agent can tell you about mail it is not going to answer. Email Manager has a new setting for what reaches you: only mail that needs a reply, which is what it did before and is still what a new install does; anything a real person wrote; or everything your ignore list did not drop. Until now, mail your agent decided needed no reply was dropped without a word, so a decision that went wrong looked exactly like an agent that had stopped working.
  • Email Replies is now called Email Manager. Answering is one of the three things it does, beside watching what arrives and filing it away, alongside answering.
  • Its settings are in six named groups instead of one long column, and every on-off setting is a switch at the end of its row instead of a tick box.
  • An email your agent is not handing you a draft for now gets two or three sentences instead of one line, and they carry the date, amount or deadline that was in the message.
  • Your ignore list and the limit on how many messages an hour your agent sends you are visible even when automatic drafts are off. Both were being applied the whole time, from boxes you could not see.
  • Times your agent writes for somebody else say which zone they are in. A reply offering "9am" left the person reading it to work out whether that meant your morning or theirs.
  • Installing an add-on no longer asks you something you have already told this agent. An answer you gave once is filled in for the next add-on that asks the same question.
  • You can stop one agent reading the web. On that agent's page, under Response quality, "Don't let this agent read the web" turns off opening pages and searching, for that agent and for every skill you install later. Email, calendar and chat keep working, and so do the services you connected yourself on the Connections tab. Worth turning on if the computer running Archie belongs to your work.
  • Fixed: every question you asked about an Outlook mailbox failed before it left Archie. Asking whether a message had arrived from somebody sent Outlook a request in a shape it rejects, and your agent then reported the mailbox as unreadable and named the address, which invited you to reconnect an account that was never broken. Watching an Outlook inbox was never affected, which is how both could be true on the same account at the same time.
  • Fixed: the list of who can message your agent could show a stranger's phone number under your own name. A Telegram account id is ten digits, the same shape as a phone number here, so it was drawn as one. The number was real, and it belonged to somebody who has never heard of Archie.
  • Fixed: your agent could be watching no inboxes at all while the panel said it was on. If you had chosen which inboxes to watch and later reconnected those accounts, the saved choice pointed at accounts that no longer existed. It watches all of them now, and says so.
  • The Activity log, on the agent's Setup tab, now says what your email watching is actually doing: which inboxes it is watching, when it is taking its first look at a mailbox and holding back everything already sitting in there, and when it has lost track of one. The first look is what catches people out. Send yourself a test message before it happens and that message is already behind, so nothing arrives and nothing is wrong.
  • Fixed: pressing Draft on an email your agent had decided needed no reply did nothing at all. It asked itself the same question again and got the same answer. Pressing Draft is your decision, so it writes the draft now.
  • Fixed: Text Replies could go quiet with nothing wrong. A sign-in that had expired, or a dropped connection, looked exactly like a text that needed no answer.
  • Fixed: a summary sent to Telegram could arrive with stray underscores through it, once summaries became long enough to run to a second line.
  • Fixed: replacing the token for a chat app your agent is already on emptied the list of who is allowed to message it, and the screen said the opposite. It now names who would have to claim the agent again, before you press save, and sends you to the claim code afterward.
  • Fixed: a question about what was said on a past call could be handed to your calendar instead of to the add-on that reads call transcripts.
  • Fixed: the video skills could not run on any computer that downloaded their tools rather than building them, which is every computer but a developer's. Video Synthesizer works again.
Archie 0.1.7August 25, 2026
  • Your agent now says when it cannot read a mailbox, instead of looking healthy and going quiet. A sign-in that has lapsed, or a permission your workplace withdrew, turns the agent's status to "Needs attention" and names the address, what the mail provider said, and where to press Reconnect.
  • Your agent also says when it wrote a reply and could not deliver it. A bot removed from a group, or one whose permission to post was taken away, used to look exactly like an agent with nothing to say.
  • Every connected email account has a Reconnect button now, not only the ones that could be given more access. An account that had stopped working looked perfect and had nothing to press, and the only way back was to remove it and add it again, which no screen said.
  • Fixed: connecting a calendar from inside a skill took away that account's email. It asked for calendar permission alone, and Outlook and Gmail both hand back exactly what was asked for, so everything else was dropped. It now asks for what the account already had as well.
  • Fixed: your agent could stop reading email entirely after you reconnected an account. Reconnecting changed the account's identity behind the scenes, so the list of inboxes to watch pointed at nothing, and the watcher ran every five minutes on no inboxes at all while the panel said it was on.
  • Fixed, on a shared agent: reconnecting a mailbox that belonged to one person handed it back to everybody, so their mail started arriving in the conversation the team shares.
  • Fixed: a saved thing named with an accent, like a subscription called "Uber" with the umlaut, could stop all email checking until Archie was restarted. Nothing was shown when it happened.
  • Fixed: the chat inside Archie could show a green Live badge and refuse every message you typed, after the agent restarted itself. It now keeps working, and an agent that stops on its own says so instead of claiming to be running.
  • Fixed: removing a Google account from an agent that also had Outlook could break Google Tasks days later, with nothing connecting the two.
  • Fixed: removing the last inbox you had chosen to watch started watching all of them, including other people's on a shared agent.
  • Fixed: a light whose bridge no longer accepted Archie's key was reported as a missing light, so you went looking for a lamp that was exactly where you left it.
  • Fixed: learning your writing style from your sent mail could, over the following weeks, run an Outlook sign-in down until it stopped working.
Archie 0.1.6August 25, 2026
  • A mailbox your agent could not read is now reported as exactly that, with what the mail provider said about it. An expired sign-in, a permission a workplace withdrew, or Outlook having a bad minute all used to arrive looking identical to a quiet morning: "the inbox is empty".
  • Connecting an email account opens the mailbox once, straight away, while you are still sitting there. If it will not open, nothing is connected and the screen says what the provider said, instead of leaving a green row that never works.
  • Your agent no longer insists it can read a mailbox that has never once opened. It used to be told to say it had access on the strength of the permissions alone, so it argued with people whose mail it genuinely could not reach.
  • An Outlook account at a company or school whose administrator has to approve Archie now names the setting to ask for, spelled the way their admin console spells it, and where it lives. Asking an IT department to "approve an app called Archie" gets a reply asking which permission and where.
  • The calendar is offered whichever email you use. That button was tied to Google being configured, so somebody connecting Outlook was never offered a calendar at all.
  • Searching the store for "outlook" finds the email and calendar add-ons. Every one of them names both providers now, and Email Manager and Calendar answer to "outlook" spoken aloud as well.
  • The store says which add-ons need an AI account of your own on the row, before you press Install, rather than refusing after you have pressed it.
  • Two guides that did not exist. One is how to update Archie. The other is connecting your email and calendar: where the button is, what each permission lets your agent do, what to send a workplace administrator when it refuses, and what to do when an account stops working.
  • The first-run screens say what Archie needs to run: this computer left awake and plugged into power, with the sleep setting named and where to change it, and a cable from your router if you can reach one. A routine set for 7:30am needs this computer awake at 7:30am.
  • "Needs attention" is a button. It was two words beside a small icon whose only explanation was a tooltip, so there was nothing to press and, on a touchscreen, nothing to read.
  • Six screens sent you to Settings to connect an email or calendar account. That is not where it is. They name the Connections tab, the section inside it, and the button, and they say to tick the permissions before you sign in.
  • Fixed: an error that said only that something had failed now says where to go and what to press. A refused key names the screen you paste it into; a lost connection names the cable to check.
  • Fixed: running out of credit on your own AI account names that company and its billing page. It used to send you to a company you may not have an account with.
Archie 0.1.5August 24, 2026
  • Reminders arrive with something to do about them. Done ticks the matching task off your list, Later offers an hour from now, this evening or tomorrow morning, and Dismiss closes it. Answering in words still works: say "did that" or "move it to four" and the buttons settle themselves.
  • Ask to be reminded and you get a reminder. It used to write a task instead, which is a row you have to go and look at and interrupts nobody. Now it sets the reminder, adds the task where the thing is also work worth writing down, and tells you it did both.
  • Cancel a reminder by what it is about ("cancel the invoice one") rather than by a number you were never shown. The task it was filed with goes with it.
  • Archie can move to another computer. Account has a Moving and backups section that writes one file holding every agent: their instructions, their skills with the answers you gave, routines with their schedules, records, memory and chat history. Nothing in it stores a location, so the same file works from a Mac to a Windows PC. No passwords or tokens travel, on purpose, and the file tells you what you will be reconnecting.
  • A connected account that nothing is using now says so, and names the add-ons that would use it. Connecting your email and getting nothing out of it was invisible before.
  • Every setup that ends in a long code now begins by saying where to stand. Telegram gets a step of its own, because reading a 46-character token off a phone and typing it into a laptop is the part that goes wrong. Addresses inside a step are links now, instead of something to retype while you are already holding two devices.
  • Archie for Business: you pick the departments a company needs and its agents arrive set up for them, instead of building each one. Agents are peers, with no head agent above them, and the company page shows who does what.
  • Archie for Business: a list belongs to the company rather than to whichever agent wrote it, so one department can hand another a list without handing over everything else it keeps.
  • Fixed, on Outlook: a reply would not send unless you had also allowed Archie to file mail. It sends either way now. The permission itself is named for what it does, because it decides both filing and whether a reply lands inside the conversation it answers, and it only ever said filing.
  • Fixed, on Outlook: an account at a company or school whose administrator has to approve Archie now says so, and says what to ask for. It used to say "try again and choose Accept" about a screen with no Accept on it.
  • Fixed: the Chat tab could open blank until you reloaded the window.
  • Fixed: a task list saved by an earlier version was invisible to the agent that kept it, which then said its tools were not working.
  • Fixed: your agent sometimes answered "I processed the request but couldn't generate a response".
  • Fixed: a moment of bad network was reported as a failure and left you to send the message again.
  • Fixed: the AI account panel showed Claude selected while saying OpenAI was connected.
  • Fixed: the strip above the message box showed the raw asterisks and link addresses of the message you were replying to.
Archie 0.1.4August 23, 2026
  • Your new agent asks what your days look like, then sets itself up from the answers. The skills and routines that fit your work are chosen and switched on while you are still talking to it, and it tells you what it added.
  • Its page opens on what arrived: each ability named, with what it does, and how to take any of them back off.
  • You can tell your agent to remove a skill or pause a routine and it does it, after one tap to confirm. Pausing keeps the routine and everything you set on it.
  • Waiting On follows up across texts as well as email. Ask by email, get the answer as a text, and it still closes the loop. On a Mac, with the text watch on.
  • Archie for Business asks about the business: what it does, who its customers are, and who else will be talking to it. It no longer asks a company owner their age.
  • Fixed: a new agent said hello twice, and then asked your name after greeting you by it.
  • Fixed: confirmations for adding or removing an add-on are tappable now, instead of asking you to type the word back.
  • Fixed: deleting an agent showed an error after the agent was already gone.
  • Fixed: the welcome cards played after every new agent instead of only the first.
Archie 0.1.3August 19, 2026
  • You can try Archie before paying for it. A new account gets free credits that run an agent on our AI account, or fourteen days on an AI account of your own.
  • Your agent answers on iMessage, on a Mac. That makes six chat apps, alongside Telegram, Discord, Slack, Signal and Matrix.
  • You can talk to your agent in Archie's own window. It is one conversation, so a question you start in your chat app can be finished here.
  • Send it a photo and it looks at the photo. It can make pictures too. Send a voice note instead of typing, and have replies read out loud.
  • Your agent can use a website itself, the same way you would, when there is no direct connection to it.
  • Outlook and Microsoft calendars work the way Google's do. Sign in once and your agent can read the mail and the diary.
  • "Remind me to pay the water bill tomorrow at 10am" now sets a reminder.
  • Your agent can learn how you write from what you have already sent, and draft in your voice.
  • The store has shelves, packs and a cart, so a set of add-ons installs in one go.
  • Archie has a face. Ember is the app's own character, and every agent gets its own version of him.
  • Dark mode is stone rather than brown, and every color in the app was measured against its background so nothing is left faint.
  • On Windows: the media tools (transcription, video) install on Windows now, and this is the first version Windows can update itself from.
  • Fixed: sixteen routines woke up every morning to work out whether it was their day. They run on the day they mean now.
  • Fixed: a restart could overwrite your agent's first message, and a long reply stopped one message short of the bottom of the window.
Archie 0.1.1July 27, 2026
  • Archie updates itself. When a new version is ready, a banner appears in the app; one click downloads it, verifies it, and restarts.
  • A much smaller download. Media tools (video, transcription) now download only when you install a skill that needs them, instead of shipping inside every copy.
  • Bug fixes and stability improvements across sign-in and the marketplace.