Aethel Release Notes

Lucin Solutions LLC · Aethel · Last updated: November 7, 2026

A running log of what changed in each Aethel release. Newest at the top. For a deeper walkthrough of any feature mentioned here, see the User Guide. For installation steps, see the Quick Start.

v2.23.1, November 7, 2026 (BETA_v2.23.1)

Web apps with full-screen task views, big Finder folders on a Mac, and saved logins all behave better.

Fixed

• Full-screen task views. On sites like ClickUp, opening a task fills the screen with its own window. That made Aethel's browser wait and then restart. Aethel now reads the open task, keeps going, and stays on the same page.
• Typing that opens a pop-up. When typing in a box opens a pop-up window, Aethel leaves the Enter key alone and knows it did, so it never thinks a search was sent when it was not.
• Big folders on a Mac. Asking Aethel to read a Finder window with a long list of files used to wait about half a minute and then give up with nothing. It now answers within about 20 seconds with what it read, and says when it stopped early. In that case it does not click a file by name, because another file with the same name could be in the part it did not read. It takes a look at the screen instead.
• Saved logins. When Aethel cannot read Chrome's list of saved logins in full, it no longer fills in the first one on the list, which could be the wrong account. It asks you to pick yours. When a Mac permission is what blocks the list, Aethel tells you which one to allow.
• Web chat scrolling. Scrolling up to read while a reply is arriving no longer pulls you back down to the bottom.
• Reliability: many smaller fixes from a full review.

v2.23.0, November 5, 2026 (BETA_v2.23.0)

A redesigned web chat where messages wait their turn, spoken replies take turns, and web apps with pop-up windows no longer stall the browser.

New

• A redesigned web chat. A cleaner look on your phone and computer: a working card that shows what Aethel is doing and which model is answering, a footer under every message with the time, where it was sent and who answered, Reply and Copy on every message, a reply quote that takes you to the message it answers, photos that fill their message and files shown as cards, and a recorder with live sound bars for voice notes.
• Send while she's answering. Messages you send while Aethel is busy wait their turn, in the order you sent them, on the web and on Telegram. On the web they sit at the end of the chat with a dashed outline, with a pencil to edit and an X to remove them. On Telegram you see that your message is next in line.
• Spoken replies take turns. A second spoken reply waits for the first to finish. If you pause Aethel or play a video or your own voice note, her voice stays quiet until you press play on one of her replies.
• "Aethel was updated." A chat page left open while Aethel updates offers a Reload button, and keeps what you were typing.
• Missing browser extension. When the Chrome extension Aethel uses to browse is removed, Aethel tells you once and the browser status says so, until it is back.

Fixed

• Web apps with pop-up windows. On sites like ClickUp, opening a window such as Share made Aethel's browser wait and then restart, losing your place. Aethel now reads the open window, keeps going, and closes it with its own Close button when it is done.
• Browser clicks. A button inside a box that hides part of it, a dialog in front of the page, and a site that hands you over to its desktop app are all handled, with a clear note when something needs you.
• Saved keys found under a related name. When a script asks for a key under a slightly different address than the one you saved, Aethel is told which saved key matches, with a reminder to check it is the right account (for example a test account versus a live one). A vault that is busy for a moment says to try again instead of saying the key is missing.
• Telegram messages interrupted by a restart. A voice note, photo or file whose download was cut short by a restart no longer gets stuck or disappears: Aethel is told it could not be downloaded and can ask you to send it again.
• Fewer wrong and duplicate problem reports. A connected service's own rules are not reported as Aethel problems, and the same problem seen again is added to its existing report instead of starting a new one.
• Model names. A model whose name ends in a word like Max is shown with its real name.
• Reliability: many smaller fixes from two full reviews.

v2.22.7, November 3, 2026 (BETA_v2.22.7)

Scheduled tasks and memory keep working when your plan's default model runs out, plain-text emails look the way you wrote them, and forms with hidden checkboxes are easier to fill.

Fixed

• Tasks and memory keep going when a model runs out. Scheduled tasks and Aethel's memory updates use your engine's default model. When that model is out of quota on your plan, Aethel now quietly runs the same work on the model you chat with, on all three engines. You only hear about it when both models are out, and the message names them and says whether your chats are affected. Before, everything stopped and you got an alert that was hard to make sense of.
• Plain-text emails keep their shape. An email Aethel writes as plain text keeps its line breaks and indents, and words in angle brackets such as "Dear <Hiring Manager>" or an address like "Jane <[email protected]>" now show up instead of disappearing. Emails written with formatting are sent as before.
• Forms with hidden checkboxes. On some websites the real checkbox is hidden inside a box. When a form fill stops there, Aethel now knows which fields were already filled and ticks the option by clicking its words. If a page redraws itself while being filled in, Aethel checks the earlier fields again instead of assuming they are still filled.
• Asking for a number of results. When Aethel's AI asked one of its tools for "20 emails" using a slightly different word for the count, the tool quietly used its default of 5. It now understands the common other names and does what was asked.

v2.22.6, November 1, 2026 (BETA_v2.22.6)

Web search works again, Aethel reads PDF links, Monday to Friday tasks run on the right days, and you can color-code your calendar events.

New

• Color-coded events. Ask Aethel to make an event Peacock blue or Flamingo pink, or to put it back to its calendar's color, and she can tell you which color an event has. Changing only the color sends your guests no update email.

Fixed

• Weekday schedules run on the right days. A task set for Monday to Friday was running Tuesday to Saturday. It now runs on the days you chose. A schedule Aethel cannot read is refused when you save it, and a task saved earlier with such a schedule now says so in the Tasks tab and on Telegram, instead of silently never running.
• Web search. Aethel's own web search had stopped answering for many people, because two of the sources it asked stopped working. It now asks a reliable source first. The AI's built-in search is still used first when it has one.
• PDF links. Aethel now reads the text of a PDF link directly, instead of getting unreadable data or opening it in the browser and losing the browser connection. A PDF that drops the browser connection is no longer blamed on your password manager.
• Saved logins with two parts. A script that needs both parts of a saved login (an API key ID and its secret, for example) now gets both, so services like Alpaca stop refusing the key.
• Connected services. When one of two connections to the same service (two Notion connectors, for example) is signed out, Aethel uses the one that works instead of stopping your scheduled task. Adding a second connector for a service you already connected now tells you so.
• Expired Codex sign-in. When Codex's saved sign-in has expired, Aethel now asks you to sign in again and your scheduled tasks wait for you, instead of showing an error it did not recognise. Setup points you to signing in again when its test run is refused, and Codex sign-ins with an API key are recognised on newer Codex versions.
• Telegram messages. Sending a Telegram message no longer fails the first time when the chat number arrives as a plain number.
• Answers that arrive. With Antigravity, an answer that was finished but never handed over is now delivered.
• Search by meaning on a busy computer. A slow memory search helper gets time to answer and is replaced quietly in the background, and Aethel tells you when search by meaning works again.
• Fewer wrong problem reports. A connected service's own rules (such as a Notion date format) and a question about an earlier report are no longer filed as Aethel problems, and Aethel tells you a problem was reported without a confusing report number.
• Privacy. Aethel's log no longer records the full address of every web search.
• Reliability: smaller fixes from a full review.

v2.22.5, October 27, 2026 (BETA_v2.22.5)

Search by meaning in your memory starts reliably on slower computers and on Windows on ARM, voice notes work on Windows on ARM, and your web chat never shows up in search engines.

Fixed

• Search by meaning starts on a slow computer. When your computer is busy right after it starts, Aethel now gives memory search by meaning up to five minutes to get going and tries once more, then keeps checking in the background, instead of switching it off for the day.
• Search by meaning stays up to date. An older copy of the search helper is brought up to the tested version by itself when Aethel starts, and one that will not load is reinstalled.
• Windows on ARM. Search by meaning now installs on Windows computers with an ARM processor, and voice notes find the right copy of Python there, so they are understood again.
• Setup on older systems. Installing Aethel's helpers no longer fails on a computer with an older Python installer.
• Your web chat stays private from search engines. Search engines are told never to list your web chat page, including when you use remote access.
• Fewer wrong problem reports. When a service you connected yourself (such as Notion) turns down a request because of its own rules, Aethel no longer reports it to the team as an Aethel problem.
• Reliability: smaller fixes from a full review.

v2.22.4, October 26, 2026 (BETA_v2.22.4)

Pictures the AI makes now reach your chat, Excel edits are back to taking seconds, and search by meaning in your memory keeps working on a busy computer.

Fixed

• Pictures reach your chat. When Codex or Antigravity makes a picture for you with its own image tool, Aethel now sends it with the reply, on Telegram and in web chat. Before, the AI could say "here it is" while nothing arrived. If the connection drops right after a picture is made, you still get the picture.
• Files Aethel names in a reply reach you. A file Aethel made during the reply now arrives even when it is written as code, and every file in a reply goes through the same safety check as files Aethel sends on purpose, so passwords and sign-in data are never sent.
• Saved web pages download. A web page Aethel saved for you now downloads when you open it, instead of opening inside the chat window.
• Faster Excel changes. An Excel change that sometimes took about a minute now takes a few seconds.
• Office's Recent list stays yours. Aethel's temporary working files no longer show up in Office's list of recent files.
• Spreadsheets made by Python. An Excel change is no longer refused when the workbook has empty cells written by Python, and a sort puts empty text cells where Excel itself puts them.
• Search by meaning stays on. One slow moment no longer switches off searching your memory by meaning; Aethel starts a fresh helper and carries on. The nightly memory update also finishes on slower computers.
• Long Antigravity jobs. When Google's sign-in for Antigravity runs out during a job of more than an hour, Aethel tells you plainly what happened and that the finished part is kept.
• Fewer wrong problem reports. A dropped internet connection inside the AI's own image tool is no longer reported to the team as an Aethel problem.
• Reliability: smaller fixes from a full review.

v2.22.3, October 20, 2026 (BETA_v2.22.3)

Searching on websites is faster, PDFs that open in the browser are read in seconds, and you can follow what Aethel does in the browser more easily.

New

• Follow along in the browser. When Aethel works on something below the part of the page you can see (a form field, a button, words she found), the page glides so that spot sits in the middle of the screen, with her gold pointer on it. Lists and menus inside the page stay where they are.

Fixed

• Search boxes. Searching in a box that changes as you type, such as Workday's, no longer waits about 20 seconds and fails. When a suggestion that matches what she typed appears, Aethel picks it instead of opening a slow results page; otherwise she presses Enter as usual.
• PDFs in a browser tab. A policy or document that opens as a PDF in the browser (SharePoint, OneDrive) is now downloaded and read in seconds instead of minutes, and Aethel tells you which folder the copy is in.
• Honest clean-up. When another app, such as Adobe Acrobat, opens a downloaded file by itself and keeps it locked, Aethel leaves it and tells you where it is, instead of saying everything was cleaned up.
• Reliability: smaller fixes from two careful reviews.

v2.22.2, October 19, 2026 (BETA_v2.22.2)

Fixed

• Usage statistics wait for setup. Aethel sends its daily usage summary only after you finish setup. Before that, nothing is sent.

v2.22.1, October 17, 2026 (BETA_v2.22.1)

Aethel now asks you one plain question before a change that needs your yes, and simply makes a change you named, with undo to put it back. Long voice replies are ready sooner and sound clearer, and voice notes that stopped working after a speech-recognition update are repaired by themselves.

New

• One plain question, never a code. When a change needs your yes, Aethel describes it in her own words and asks once; you just say yes. You never see or type a code.
• A change you named is just made. Deleting the rows, column, tab, slide or paragraphs you named, or sorting a table, now happens straight away in Google Sheets, Docs, Slides and Forms and in your Excel, Word and PowerPoint files. Undo puts it back. She still asks first when the change would lose something undo cannot bring back, such as a chart, a comment or a Google Table.
• The Trash when you ask. A Drive file or folder you ask her to move to the Trash goes there at once, and stays in Drive's Trash for 30 days.
• Voice replies, faster and clearer. A long spoken reply is made in parts at the same time and joined only between paragraphs, so it is ready sooner and plays without gaps, in higher sound quality. Dates such as "1966. godine" are read the way they are written, without a pause in the middle.
• News just for you. The Aethel team can now send a product note to your installation alone, for example the answer to a question you asked. It arrives like any other Aethel news, and you can turn news off in Settings.

Fixed

• Voice notes stopped being understood on some computers after a speech-recognition library update. Aethel now repairs that by herself in the background when she starts, and tries again every hour if it did not work. A voice note she still cannot understand gets a real answer telling you why.
• Excel workbooks built for you with a sheet name Excel does not allow (more than 31 characters, for example) are now caught: Aethel names the sheet and rebuilds the file, instead of handing you a workbook Excel offers to repair.
• Saying yes to a picture for a shared document or a large spreadsheet goes through the first time.
• Deleting a sheet tab that holds a Google Table now asks first, because undo cannot bring the Table back.
• Reliability: many smaller fixes from two careful reviews.

v2.22.0, October 15, 2026 (BETA_v2.22.0)

Aethel can now keep a whole folder copied to Google Drive in the background, sign you in to websites with the logins Chrome already saved, and set up and look after connectors to your other apps on her own, with their keys kept in your Vault. Voice replies work again after Microsoft changed its voice service, and problem reports she sends are matched to ones already reported.

New

Google Drive

• Keep a folder copied to Drive. Ask "keep my Invoices folder copied to my Drive" and Aethel copies it in the background: only new and changed files go up, the newest copy wins, and nothing on Drive is ever deleted. Ask how far along it is or tell her to stop at any time; stopping never leaves a half-sent file. "Preview the sync" shows what it would do without changing anything. A file that stores sign-ins or keys is skipped and named, and a name she cannot match on Drive (two folders with the same name, say) is reported, and that part is never called done.

Browser

• Your saved Chrome logins. When a site needs you signed in, Aethel picks the login Chrome already saved for it, the way you would. If Chrome holds more than one login for the site, she asks which account to use, and she always tells you which account she signed in with. She never clicks or types while another window is in front of Chrome.

Connectors

• Finding a connector. "What are my options for Asana?" now shows three kinds of options together: a ready-made connector, one published elsewhere, and the service's own API.
• Keys in your Vault. A key a connector needs is kept in your encrypted Vault instead of the settings file.
• Signing a connector in. For a connector that signs in with an online account, Aethel opens its sign-in page in Chrome and clicks Allow for you, each time only after you say yes. Signing in again keeps the sign-in you had if the new one is refused.
• Kept up to date. Once a week Aethel tries each connector's newest version and keeps it only when it works at least as well. An update that would remove a tool you used in the last month is held back, and a connector whose key stopped working asks you for a new key.

AI engines

• The engine's own web search first. Each AI engine now searches the web with its own search, which is faster and more reliable; Aethel's search is the backup, with a second route when the first is busy, and says plainly when no search could run.
• Model lists from the engine. The Model menu in Settings lists the models the engine itself offers your account, refreshed in the background, and the sign-in state is the engine's own answer.

Problem reports

• Already reported? Before Aethel files a problem, she checks whether it was already reported; if so, she adds yours to it and can tell you what the team did about it. Screen recordings and phone photos can now be attached, and she only offers files a report can actually carry.

Fixed

• Voice replies work again after Microsoft changed how its voice service signs in, and a slow connection no longer delays a long reply.
• Emails to other people are shown first. In a chat, Aethel shows you an email or message to someone else and sends it after your yes, unless you said "just send it".
• A retired Codex model in Codex's own settings no longer stops your chats: Aethel runs on the engine's current default and carries on. Settings and the Dashboard name the model Aethel actually runs, and switching engines from a chat updates an open Settings window.
• Engine sign-in: Settings never calls an engine signed out on a guess, and Sign in again shows after an engine refused a run.
• Browser: a click Chrome never passed to the page, a link that opened an app, a page that moved to a new tab, and a tab that was already closed are each explained in plain words, and no stray Chrome "Welcome" tab is left open.
• Connectors stay signed in when their sign-in helper updates, a connector that stops working is checked again, and its key never appears in a log.
• Files holding passwords or keys are never sent by chat, email, Drive or a website upload without your yes, and that yes covers that one send only.
• Large Drive uploads no longer stop part way.
• Downloads watcher: a folder name like "Bank" somewhere above your Downloads folder no longer hides every download.
• Reliability: a problem report whose connection drops is sent again, once, and many smaller fixes from a careful review.

v2.21.5, October 12, 2026 (BETA_v2.21.5)

Aethel now changes your Excel, Word and PowerPoint files themselves instead of leaving a trail of copies, an email she writes for you goes out under your own name and quotes the original the way Gmail does, and a chat no longer stops when the AI engine's standard model is retired. Also in this release: scheduled tasks that tell you when a step was skipped, and a round of fixes from a careful review.

Fixed

Office files

• The file itself is changed. A change to an Excel, Word or PowerPoint file is now made in that file, wherever it lives, with the same checks and the same undo as before. Copies with "(Aethel)" in the name are made only when you ask for one: "save it as a new file" makes a copy for that one change, and "make copies from now on" keeps copies for that file until you say otherwise. A scheduled task that names the file switches copies off again, so a daily update never stops because copies were on.
• A read-only file is not touched. Aethel asks whether to make it writable or to put the change in a new file; a scheduled task flags it for you. A refusal that is not about read-only is named for what it is: the file's or folder's permissions, or a program holding the file.
• A file open somewhere is left alone. A workbook open in Excel on this or another computer is never written and never copied; Aethel asks you to save and close it first. An undo refused for the same reason says it was not done.
• A big file in a scheduled task. A task that names a file over the undo limit (8 MB) changes it and tells you the change cannot be undone, instead of skipping it every day.
• Undo picks the right change. "Undo that" after a change to a file followed by a copy of it undoes the newer of the two.
• On a Mac, when the Office app cannot say which files it has open, Aethel tells you why (an Automation permission, or a dialog waiting in the app) and what to do, and never launches the app to find out.

Email

• Your emails show as yours. An email Aethel sends for you goes out under your own Gmail name, and Gmail no longer marks your own copy "not authenticated". A name with accents or a comma keeps its address.
• Replies quote like Gmail. A reply now quotes the original with its formatting, links and signature pictures, the way Gmail does, instead of plain text with broken line ends. Anything in the original that could load or run something is dropped from the quote, and a quote that cannot be carried (a picture that cannot be fetched, a very large original) falls back to a text quote rather than a failed reply.
• A momentary failure looking up your sender name no longer leaves every later email without it until Aethel restarts.

AI engine

• A chat never dies because a standard model was retired. When the engine maker retires the model Aethel uses by default, she switches to the engine's own default and carries on. A model you picked yourself is still reported to you as before.

Scheduled tasks

• A skipped step reaches you. When a task skips a step (a file it may not change, a question only you can answer), it says so in the message it sends you, or Aethel texts you, instead of noting it where nobody reads.

Browser

• A click you may have made yourself. When a link Aethel clicked lands somewhere else while you are using the computer, she says you may have moved the page rather than reporting a fault; a scheduled task opens the link's address and carries on.

v2.21.4, October 12, 2026 (BETA_v2.21.4)

Aethel works better beside you in the browser: when you move a page she is using, she knows it was you, and a link that does nothing when clicked no longer leaves her stuck. Voice replies that the voice service cuts off part way are now finished, never sent half done.

Fixed

Browser

• A page you move is yours. If you go to another page in a tab Aethel is working in, she now sees that you did it, not that something went wrong. A scheduled task takes the page back to where it was and carries on; in a chat she asks you first. While you are at the computer, a tab she needs that is behind another one comes to the front straight away instead of after a wait.
• No double clicks. When a click seems to get no answer, Aethel tries it once more only if the same button is still on the same page. A click that only puts the cursor in a box now tells her what is on the screen, so she does not keep clicking.
• A link that goes nowhere is named. Some links, on LinkedIn for example, do nothing when clicked. Aethel now sees where the link points and can open that address herself. A link that opens a new tab says which tab it opened.

Voice replies

• A cut-off voice reply is finished. The voice service sometimes stops sending part way through a reply. Aethel now notices, fetches the rest, and plays and saves the whole reply, never just the first part.

v2.21.3, October 10, 2026 (BETA_v2.21.3)

Aethel's daily update of your AI engine (Codex, Claude Code or Antigravity) is now careful never to leave it unable to start: she checks each new version works on your computer before keeping it, and goes back to the version that worked when it does not. Also in this release: calendar moves that keep each time zone, and Excel requests read more accurately.

Fixed

AI engine updates

• A new engine version that is not ready for your computer is left alone. Engine makers sometimes publish a new version a few minutes before the part that runs on your kind of computer. An update in that gap used to leave the engine unable to start, and the next task failed with an error Aethel did not recognise. She now checks the version is complete for your computer first, quietly waits when it is not, and tries again later.
• Every update is checked. After updating, Aethel makes sure the engine actually starts. If it does not, she puts back the version that worked, on her own, and tells you only if she cannot.
• No reinstall when nothing changed. The daily check used to reinstall Codex and Claude Code even when they were already on the newest version. Now an engine that is current and working is left alone.
• A missing Antigravity is put back. If Antigravity is the engine you use and it disappears from the computer, Aethel reinstalls it herself instead of only telling you.
• A half-installed engine is named plainly. When the engine is only partly installed, Aethel says so and how to fix it, instead of reporting an error she does not recognise.

Calendar

• A trip keeps both of its time zones. Moving a New York to London flight to another day keeps its arrival in London time.
• Moving a meeting to another calendar and another day at once keeps it at the same hour, also when the two calendars are in different time zones.

Office files

• Words you want typed into a cell stay words. Asking Aethel to write a sentence like "please edit it directly" into a cell no longer changes how she saves that file.
• A greeting no longer gets in the way. "Hi Aethel: please update Cjenik.xlsx directly" is now understood as the request it is, without an extra question.

v2.21.2, October 9, 2026 (BETA_v2.21.2)

Aethel can now edit your own Excel file directly when you ask her to, and change many rows in one step. In the browser, a scheduled task and a chat no longer get in each other's way, a click on a Chrome tab that is not on screen is made by Aethel herself, and the gold cursor shows every step she takes. Also in this release: calendar moves that keep their time, web pages with the right letters, and a long list of fixes from three careful reviews.

New

Office files

• Many rows in one step. "Set these 25 products to In stock and change their prices" is now one checked change instead of 25: she finds each row by its words, makes every change, reads the file back to prove each one, and one "undo that" takes them all back. She can also add a new sheet already filled with rows.
• Edit your own file directly. Aethel normally saves her changes to a new file next to yours, <name> (Aethel).xlsx. Say "edit Cjenik.xlsx directly from now on" and she changes that file itself instead, checked the same way, and "undo that" still puts it back for 48 hours. The file keeps its own permissions (and its tags on a Mac). Say "make copies again" to go back. A sentence about something inside the file, like "change the original title" or "status: no copies", never switches it; when she is not sure, she asks first.

Browser

• The gold cursor shows every step. Waiting for a page, scrolling a list, going back, switching tabs, answering a pop-up and reading one part of a page now each show where Aethel is and what she is doing, never moving the page itself.
• A task and a chat take turns. When a scheduled task is using the browser and you ask Aethel for something that also needs it, whoever started first keeps it. The other waits up to half a minute (a chat) or a minute (a task), then is told the browser is busy, and nothing is done on the wrong page. If a task used the browser between two of your messages, Aethel looks at the page again before she clicks.

Problem reports

• Each report says who filed it: you, the assistant on its own, or a scheduled task.

Fixed

Browser

• A click on a tab you cannot see. When Chrome is behind other windows, minimized or on another desktop, a click used to wait 20 seconds and then ask the assistant to try again. Aethel now brings Chrome to the front and makes the click herself. In a chat the cursor notices sooner, so Chrome comes forward before the click and there is no wait at all. Decorative cursor moves are skipped while Chrome is not showing the page, so the other steps are not slowed down either.
• A dropped browser connection is named as one. When the connection to Chrome drops and comes back on its welcome page, Aethel now says so instead of saying a button is gone, and her log records why, including whether your tab was still open.
• No cursor on the connection page, and never a second cursor inside an embedded part of a page.

Email

• Reading a whole conversation from one email works. Asked to read a conversation using the ID of a single email in it, Aethel got "not found" and had to try again. She now reads the conversation that email belongs to.

Calendar

• A time on an all-day event makes it a timed event instead of an error from Google.
• Moving a meeting to another day keeps its time, also across a clock change (late October into November), and keeps the meeting's own time zone.
• Making an all-day event timed with only a date now asks you for the time and changes nothing until you give one, also when the event is being moved to another calendar.

Web pages

• Pages read with the right letters. A page that does not say which text encoding it uses came back with broken punctuation ("weâll" for "we'll"); it now reads the way a browser shows it.

Everything else

• A misspelled setting is named. When the assistant asks a tool for something with a setting the tool does not have, the answer now says which one was ignored instead of quietly dropping it.
• Scheduled tasks: the log says a task started only when it really starts, and a task waiting for another one says what it is waiting for.
• Undo on a Mac keeps a file private. Undoing or redoing a direct edit keeps the file's permissions and tags.

v2.21.1, October 9, 2026 (BETA_v2.21.1)

The web chat can now search your conversation, and clicking through long web pages works again on sites like LinkedIn. Also in this release: folder names with accented letters stay exactly as they are, a dropped internet connection is named as one, and a long list of smaller fixes from two careful reviews.

New

Web chat

• Find in this chat. A search button at the top of the web chat opens a panel that searches every message on the page as you type. Any word you type finds a message, case and accents do not matter ("sime" finds "Šime"), and words in quotes find that exact phrase. Results are listed newest first under Today and Yesterday headings, each with who said it and when; a message that has only some of your words says so ("1 of 2 words"). You can narrow it to your messages or Aethel's, to one kind (replies, reminders, nudges and the rest) and to today, the last 7 or 30 days or one day. Pick a result and the chat scrolls to it with the words highlighted; Enter and the arrow keys step through the rest. Ctrl+F (Cmd+F on a Mac) opens it on any keyboard, Russian and Greek included. On a phone it fills the screen, and tapping a result takes you straight to the message. Chinese, Japanese and Thai, which put no spaces between words, are found anywhere in a sentence.
• Jump to latest. Whenever you are a screen or more above the newest message, a small "Jump to latest" button takes you back down. Closing the search panel keeps the chat where you left it.

Fixed

Browser

• "Load more" buttons work again. On LinkedIn Messaging, nearly every click on "Load more conversations" failed with a message that the button could not be found. The gold cursor was scrolling the page to the button just before the click, the list loaded more on that scroll and redrew the button, and the click then aimed at the old one. The cursor no longer moves the page; it follows the scroll the click itself makes. If a button really disappears, Aethel says so plainly, and words the page shows are quoted as the page's, never as Aethel's own.
• The "Clicked" label shows where the click happened. After a click on something below the visible part of a page, the label on the next page used to appear where the previous action had been. It now never appears in the wrong place.

Files

• Folder names with accented letters stay intact. Asked to save a PDF into a folder like "Ponude i narudžbe" on a network drive, Aethel could create a broken new folder named "Ponude i narud" instead. Her tools showed accented letters to the AI as codes such as "\u017e", and the AI copied that code into the path. She now sees and writes the real letters, and the advice she follows when writing scripts no longer puts such codes in a file or folder name. Hidden characters that you would never see in a name still reach her as visible codes, so nobody can slip instructions into a file name.
• Looking up what Aethel did finds accented words. Asking what she did with "narudžbe" now finds the step that used it.

AI engines

• A dropped internet connection is named as one. When your internet is down, Codex answers with "workspace routing discovery failed". Aethel used to call that "an error I don't recognise"; she now tells you to check your connection, and no problem report is sent for it.

Problem reports

• Suggestions arrive as suggestions. When the assistant sends the team an idea for something that worked badly, it now reaches the developers as a suggestion instead of as a problem that blocked your task.

Behind the scenes

• The developers' dashboard is easier to read on a phone, notices a new version right after you update, and shows only the problems the team can act on, with one rule for when an install is not updating.

v2.21.0, October 8, 2026 (BETA_v2.21.0)

Aethel now reports her own problems. When one of her parts fails, she sends the developers a problem report by herself, so it gets fixed without you having to notice it or report it. The AI assistant does the same when it runs into something broken, or something that works badly, while helping you. Your files are never sent unless you say yes in the chat. The privacy page describes what a report carries. Also in this release: two things seen on a Mac, a click in a browser tab that was not on screen, leftover connection tabs that stayed open when many tabs were, and the Antigravity engine running without a safety check it has on Windows.

New

Problem reports

• Aethel reports her own failures. When the search component stops, an update cannot be installed, remote access keeps dropping, the chat cannot be saved, the nightly maintenance fails with an error of her own, the AI engine returns an error she does not recognise, or Aethel crashes, she sends the developers a report by herself: her own log, her settings with secrets removed, and the facts about the computer. At most once a day for each kind of failure, never for a problem on the provider's side or on yours (a used-up plan, a sign-in that expired, a dropped connection). You see one short notice in the chat the first time, and nothing after that.
• The assistant reports as it works. When the AI hits a defect in Aethel it reports it right away and tells you so ("I reported it (#141)"); it no longer asks whether it may. When it notices something that works badly, a step that takes too many tries or a tool answer that is confusing, it sends the team a suggestion the same way. The only question it asks is whether it may add one of your files, named, and only after you say yes.
• One report per problem. When one of Aethel's tools fails while she is helping you, she reports it by herself. If the assistant then reports the same failure, what it wrote is added to that same report instead of starting a second one. A different problem in the same tool is still reported on its own.
• Your files go once, and you hear where they went. When you agree to add a file to a report, it is added once: asking again, or two reports in the same chat, never uploads it twice or loses your yes, and Aethel tells you which files reached the report and which did not. A crash report's title never contains the name of one of your folders.
• One place for every problem. The developers' dashboard lost its separate health page; every failure, reported by Aethel or by the assistant, is one issue with everything needed to fix it, and the fix closes it.

Usage statistics

• Your new version is known sooner. After an update, Aethel sends her short usage summary once more right away, so the developers see which version you run within minutes instead of the next day. It carries the same information as the daily one.

Fixed

Browser

• Leftover connection tabs close even with many tabs open. Aethel closes the "Welcome" tabs left behind when the browser connection had to be set up again. With many tabs open, Chrome hides the close button on every tab you are not looking at, so these tabs used to stay. They now close, and the tab you are on stays in front.
• A click in a tab Chrome is not showing is named, not guessed at. When Aethel's tab is behind another tab, in a minimized window or on a locked screen, Chrome draws nothing for it and a click, hover or drag never starts. That used to come back after 20 seconds as "the page was slow, the action may still have been applied". Now the gold cursor notices the tab is hidden before the click, Aethel brings her tab to the front once (a minimized window is restored on an unlocked desktop), and the click runs. If the tab cannot be shown, the answer says exactly that, what still works, and that nothing was clicked. A scheduled task reports the step as blocked rather than guessing.
• "You seem to be at the computer" is said only when it is you. Windows and macOS count Aethel's own typing and clicks as yours; that is now taken out, so a click right after she pasted into an app no longer claims you were using the computer, even when two of her actions overlap.
• A slow or covered button keeps its real reason. A timeout caused by a covered element is reported as that, instead of a retry on a hidden tab that would cost another 20 seconds. A screenshot or a scroll that times out after a retry also says why, instead of "may have been applied".

Antigravity engine on a Mac

• The engine's own checks are back on a Mac. Antigravity records its working folder under a different spelling of the same path on macOS, so Aethel never found its transcript there. That quietly switched off the check that her instructions arrived whole, the live tool chips in the chat, the activity signal that keeps a long run alive, and the stop on a spent plan quota. Every one of these now works on a Mac as it does on Windows.

v2.20.1, October 6, 2026 (BETA_v2.20.1)

A small release that follows the day-after review of v2.20.0. Every finding was fixed, most of them in the browser, where a few wasted steps on a web page have been removed, and in the problem reports we receive, which now say more about a lost browser connection.

Fixed

Browser

• Clicking an element she just read works the first time. When Aethel reads a web page she gets a short code for each button, link and field. Written the way the page listing prints it, that code used to be rejected as "does not match any elements", so she fell back to running scripts in the page to find the element again. The code is now accepted in any spelling, on a click, a hover, typing, a dropdown, a drag or a form fill.
• Reading one part of a page gives a better hint. When Aethel asks for part of a page by its heading or role and nothing matches, the note now shows her the spelling that does match (an exact quoted heading, for instance) instead of one that could never work. A read of one part also says plainly that the codes from the rest of the page stop working until the next full read, so a "not found" no longer costs a guess.

Behind the scenes

• A lost browser connection says why, and who was driving. A problem report now says whether the browser extension never joined or the reconnect failed, and whether a chat or a scheduled task was using the browser at the time.
• Fewer phantom errors from our own test machines, so the picture we see of real installs is cleaner.

v2.20.0, October 5, 2026 (BETA_v2.20.0)

A release about documents that look good, not only correct. When Aethel builds a new presentation or Word file she now aims for a design you could send to a client, turns every slide or page into a picture, looks at each one and measures the layout before she says it is done. The same eye reaches Google Slides and Docs (pictures placed by words, tables, fonts and colours), the calendar warns you about clashes before you book, and the browser and the desktop show exactly what she is doing.

What's new

Excel, Word and PowerPoint on your computer

• A new deck or document is built to look good, then checked. "Make a 3-slide proposal from my template, one hood per slide with a big photo and the price": she builds it from your template with full control over every photo, box and size, and the checks (no squeezed photo, no text past its box, the file opens in the app) are a floor she must pass, not the goal. The answer ends with the designer's question: is this the look you want, or should she take it further?
• She looks at the result. "Render" turns each slide or page into a picture she looks at one by one, on Windows and on a Mac, and measures where every text and picture is drawn: the free space on a slide, a photo's share of it, a caption sitting on a frame's edge, a photo stored turned, text running out of the top of its box. A second render says per slide what changed since the last one.
• Reading a deck says where everything is. "What is on slide 3?" names each shape's box in points and inches, its font, whether the text shrinks to fit, and which layout boxes it uses.
• Bring a slide from another deck. "Bring the 'How we work' slide over from last year's proposal" copies it in, keeping its slide number and date, without a second copy of your template's logo.
• Photos beside text. "Put this photo next to the text on slide 2" places it beside the words and makes room; photos keep their shape, phone photos show at the width you asked for on both systems, and photos the app would convert on the way in (WebP, HEIC, AVIF, SVG, BMP) are refused up front rather than silently changed.
• Word tables and lists in one go. "Add a table with these rows after the price paragraph" builds a checked table; a list of paragraphs goes in as one change.
• Text that no longer fits is shrunk, and you are told by how much. Replacing words or adding a slide on Windows fits the text to its box and says how much smaller it became.

Google Docs, Slides and Drive

• Fonts and colours by name. "Make the title 18 point, dark blue, in Georgia" sets size, colour and font family; a colour with a link or a heading in the same request comes out as asked.
• A table as one checked, undoable change. "Insert a 3-column table after the budget line", and a read with positions shows each cell so a follow-up can fill them. A Markdown table in a new Doc becomes a real table.
• Pictures on a slide placed by words. "Put the chart under the title", "move the logo to the bottom right", "make the photo half as wide": pictures are placed, moved and resized by what you say, never over the title, and the read says where every element is drawn, in points.
• Replace a Google file's contents in place. "Replace the contents of my 'Offer 2026' Doc with this file" swaps the words and keeps the link, the sharing and the comments; a Sheet is not replaced this way, and a Doc or deck someone else owns is shown first.
• Text that outgrows its box is caught even over empty space, and a rewrite that makes overflowing text fit is kept rather than put back.

Calendar

• A booking warns about clashes. Creating an event or moving one now reports the events it overlaps, across every calendar you keep, including a repeating series ahead. More than five clashes says "more than 5" rather than listing exactly five.

Browser and desktop

• The gold cursor shows everything she does. It follows every click, field, submit and lookup in a script, points at what she waited for and at a page she opened, and nothing scrolls out from under you. A click that opened a new page says "Clicked X".
• Lists that load as you scroll. "Scroll to the end of the results" scrolls step by step, reads the new items, and says when the list is still growing or when it stopped.
• Desktop controls show what they hold. When she reads a window's elements, each dropdown, field and checkbox line ends with its current value, so a half-filled form is read the way you see it. A password field is never read.
• Leftover "Welcome" tabs from her browser connection are closed on Windows, only the ones she left behind, never yours and never her live tab.
• A page that cannot be reached is said in plain words: she tells you and stops, instead of hunting for it.

How she answers

• She does what you asked, then names the better option. If she noticed a risk or a better way while doing it, she says so after the result, once, instead of asking first or doing something else.
• Live facts are looked up before they are stated. A date, a price, a status or "what is on today" is read fresh rather than repeated from memory.
• A nearby search that hits a busy map service says so, falls back to a second map, and never passes old places off as current. "Near me" uses your last shared spot; "near the office" is looked up by name.
• Tasks are warned when their instructions copy a note: a scheduled task that pastes a memory note into its text is told to read the live note instead, so the task does not go stale.
• When something is Aethel's own fault, she says so and offers to file a problem report, instead of blaming the request.

Fixed

Office and Google files

• A Mac render in a synced folder leaves no stray temporary PDF, and a render that stopped part way says which pages are missing and why.
• A card, a band or a play button on a photo is no longer called its frame, and a caption with a normal margin inside a card no longer "sits on the edge".
• A photo shown upright by a flip and a turn is no longer called turned; a photo-heavy deck keeps its slide positions and flags; a slide carrying the template's own logo can be copied in.
• Photos beside a growing text box are not refused, and undoing them puts the text box back.
• A Mac sorts text the way Excel for Mac does, spaces and accents counted.
• A Word batch finds its anchor once, undoes identical lines correctly, and refuses a table touching a table.
• A Google Doc or deck too big to export is still replaced, and a lost answer from Google is settled by the file's own modified time, never by a guess.
• A converted Word document's undo works; a row added above merged totals goes in right; a cell write names its row by the words that found it.
• A sheet undo or redo never takes someone else's note, row or redo, and a rate limit on the change stamp never makes a change impossible to undo.
• A web photo over 5 MB still goes on the slide, a too-big one is said to be too big, an animated GIF keeps moving, and phone photos no longer fill Aethel's memory.
• A turned or mirrored element on a slide is read and placed where it is really drawn.

Calendar and tasks

• A time with a zone is read in the sender's zone, and a timestamp used as a range bound is read as its day, never as an empty calendar.
• A Google Task due date with a time of day no longer makes Google refuse it.
• A task that did not run while the computer slept is caught up on its own day only, once, never the next day's run hours early, and a catch-up left with no runner is picked up by the next tick.

Browser, desktop and engines

• A navigate that downloaded a file names the file, and a page quoting an error is still a page.
• A typed message with bullets or a quote is hidden whole in any error text or log.
• A pointer that did not land is explained by what was measured (a window in front, an edge, a remote desktop), never always as "the system discarded it".
• Antigravity's dropped connections read as a network problem, never as "verify your account".
• Codex's 5-hour limit that resets after midnight is no longer called the weekly one, and the usage-limit message follows the reset Codex stated, including "Oct 6th, 2026 5:00 PM".
• A sign-out only the reply mentions is not trusted; a real sign-in wall from the engine itself is.
• The memory summary pauses at a usage wall like tasks do, for the whole engine until the reset, a sign-in or an engine change, instead of retrying every 15 minutes per chat.
• A Chrome window with no readable tab strip is unknown, never read wrong.

Behind the scenes

• Problem reports ask one file question per report, never the same one twice, and your one yes lets the file Aethel asked about go with it.
• Aethel's health reporting says where each error happened (a chat, a task, a summary, a nightly run) and when a timeout was the computer sleeping, so the team fixes what is ours and leaves what is not.
• A memory you asked her to keep stays in her always-shown facts, your own name first, a merged memory keeps its place, and "facts left out" counts every one.
• Every nightly maintenance run reads only the accounts of the user it is working for, and its audit credits a read exactly as it happened.
• The build checks that every shipped library matches on Mac and Windows, and refuses a release that would wrap an older installer.

v2.19.0, October 1, 2026 (BETA_v2.19.0)

A release about changes you can trust. When Aethel edits a Google Sheet, Doc, Slides deck or Form, or an Excel, Word or PowerPoint file on your computer, she now finds the place by its words, makes exactly the change you asked for, checks the result herself, and can undo it. She also remembers what her own actions returned, so a follow-up like "send that file to Jane" finds the right file, and she works the browser and your desktop in far fewer steps.

What's new

Google Sheets, Docs, Slides and Forms

• Edits are checked, and every one can be undone. "Add a row above the Total", "replace Kickoff with Launch in the Intro tab", "change the title on slide 3", "make the dietary question required": she finds the row, paragraph, slide or question by what it says, never by a number she guessed, and when several match she lists them and asks. "Undo that" puts back her last 10 changes per file from the last 48 hours, and is refused, changing nothing, if someone has edited that part since.
• Dates stay dates and formulas stay formulas. A sheet edit no longer turns dates or numbers into text, and a Total keeps adding up after a row is added.
• Docs keep their look. A replaced bold word stays bold, a new paragraph matches its neighbours, and a change inside someone's pending suggestion is left alone.
• Slides that overflow are put back. After a text change she compares the slide before and after, and if the text spills out of its box she undoes it and tells you.
• Form answers are never lost by accident. Removing a question that already has answers is shown first, and undo brings the question and its answers back.
• Moving, sharing and trashing files is checked too, with a preview before anything that lets someone new open a file, and undo for each.

Excel, Word and PowerPoint on your computer

• Named changes, checked after saving. Add a row, change a cell, sort, replace words (headers and footers too), add a paragraph, add, copy, move or delete a slide, swap the logo on every slide, export to PDF. After the app saves, she reads the file back and keeps it only when it holds exactly the change you asked for.
• She can look before she changes. "What is in this price list?" reads the sheets, rows, paragraphs and slides straight from the file.
• Undo for Office files, the last 10 changes per file for 48 hours, kept on your computer only.
• PDF to an editable Word file, through Word itself on Windows, and on a Mac through Word's online conversion after you allow it once.
• Every result says what it holds and which pictures are new, and on Windows a file built without the app is opened in it to prove it will open.

Aethel remembers what she did

• "3 tools used" under a reply. In the web chat, any answer that took actions shows a small line you can open to see what each step did. Aethel sees the same lines in later messages, so she no longer has to look things up again or guess which file, event or email "that one" was. Codes and passwords are hidden from them.
• Older parts of a long conversation stay within reach. Aethel carries short summaries of the messages she can no longer see, and can search the whole conversation word for word when you ask about something said earlier. A quiet line in the web chat shows where her view begins.

Faster browser and desktop control

• Fewer steps for web tasks. She fills a whole form, clicks and presses keys in one go, waits for each page to settle before she looks, and after an action reads only what changed on the page. In our tests five common web tasks took about a third less time and half the steps.
• She can close the tabs she opened and leave yours alone.
• Clicking a button by its name in a desktop app, with a check that the right app is still in front before each click or keystroke.
• Typing into apps is a paste that puts your clipboard back. Windows 11 Notepad and some other apps scrambled typed text, so Aethel now pastes on both Windows and Mac. Everything you had on your clipboard, a picture included, is set aside and put back once the app has read the text. In remote desktop and virtual machine windows she still types key by key.

Web chat

• No Telegram bot needed. If you never set one up, reminders, nudges and notices arrive in the web chat, and wait there until you open it.
• Slash commands and their answers show on every tab, from Telegram and the web alike, and a failed reply keeps its Error mark after a reload.
• A new header. Your name and initial, a pin and Sign out sit at the top right. Your address no longer shows on screen all the time; it opens only when you click the pin. Sign out asks before signing you out.

Fixed

Conversations

• Two messages sent close together are each answered in their own turn, and Aethel no longer loses her own answer when a second message arrives before it.
• An open web chat tab and a reload now show the same conversation, including messages that arrived on Telegram while the tab was open.
• "Delete all memory" stays deleted, even when a background summary is being written at the same moment, and a fact you asked her to remember is never folded away.
• A long conversation never loses messages nobody summarized yet.
• A Telegram reply is saved as delivered only when it reached you, and files sent with it stay listed on it.

Office and Google files

• On a Mac, an Office job can no longer save over the workbook you have open. A job now works only on the document it opened by name.
• Word's Track Changes being on is now mentioned, so you know the edits arrived as revisions.
• A document the job left open is closed, and a damaged file that hangs while opening is named.
• Excel sorts text the way your Mac or PC sorts it, emoji and rare characters included.

Browser, desktop and engines

• A typed password never comes back in what the browser reads from the page.
• A program that jumps in front during a click by name no longer gets the click.
• One stuck connector no longer delays every Codex reply and task by up to two minutes.
• Codex retrying a connection is no longer reported as the turn failing.
• A sign-in that expired on Claude is reported as a sign-in problem, not a generic error.

Behind the scenes

• Aethel's own records are read-only to her tools. Her Python and file tools can read your chat history, her undo copies, settings and log, but can no longer change, move or delete them, and her Python scripts cannot read your saved sign-ins at all.
• Problem reports ask before sending your documents. A report that mentions a file the team cannot see asks you once whether to attach it, and an Office or PDF file is attached only after your yes.
• The crash log is kept to a reasonable size by the daily cleanup.

v2.18.1, September 29, 2026 (BETA_v2.18.1)

A release about email that behaves the way your mail app does. Replying, forwarding and drafting through Aethel now follow the same rules Gmail follows: a reply goes where the sender asked, a reply to your own mail goes to the people you wrote to, a forward is one message, and a reply Aethel drafts for you lands in the right conversation. A tester's report drove a full check of every email feature, and this is the result.

What's new

Email

• Add someone back onto a reply. When an earlier sender hit Reply instead of Reply All and dropped a person, you can say "reply and put Jane back on copy". Aethel adds them to the threaded reply instead of forwarding the message and starting a new conversation.
• A drafted reply stays in its conversation. "Draft a reply to Jane for me to review" now saves a real reply draft, in the same Gmail thread with the quoted original and the "Re:" subject, so when you press Send in Gmail it goes out as part of the conversation rather than as a new one.
• Archive, star and spam in one step. "Archive those", "star this one", "mark that as spam" and their opposites are direct actions now, on single messages or whole conversations. Marking mail as spam is for the admin only, since Gmail deletes spam after 30 days, the same as the Trash.
• Forward with your own files. A forward can carry your own attachments, an inline picture and a priority flag, the same way a new email can.
• A long search is not cut short. An email search past a hundred results used to stop there without saying so. Aethel now tells you there is more and can read the next page.

Fixed

Email

• A reply goes where the sender asked. A message from an address like noreply@ with a "reply to" of support@ was answered to the dead noreply@ box. A reply now goes to the address the sender named. A reply-all still keeps the sender on copy, the way Gmail does.
• Replying to your own sent mail works. It used to fail with "no reply recipient could be derived". A reply to a message you sent now goes to the people you wrote to, and a note to yourself replies to you. Mail sent from one of your other Gmail addresses is recognised as yours too.
• A forward is one message. Forwarding to three people with someone on copy used to send three separate copies, so the copied person got three and the recipients could not see each other. It is now one message, and the reply that names it is right.
• Recipient lists separated with semicolons work everywhere. [email protected]; [email protected] used to be read as nobody: a forward refused, a new email went out with an empty To line, and a draft was saved with no recipients. Send, draft, reply and forward all accept a comma or a semicolon now.
• A reply aimed at the wrong end of a conversation is stopped before it is sent. When the target turned out to be the first message of a longer thread, Aethel used to send the reply to that oldest message's recipients and only then mention the mistake. She now refuses before sending and tells you how to reply to the latest message.
• Names with accents show as names in the confirmation of a reply, draft or forward, instead of as encoded text.
• On a two-account setup, the next page of a search is read from the same mailbox it started in.

Behind the scenes

• Less noise in the log. A harmless notice the Antigravity engine sends at every start is written once per session instead of twice per call (one report held 197 of them), and the log respects your log level for these lines.

v2.18.0, September 26, 2026 (BETA_v2.18.0)

A release about your Google Drive, and about replies that look right wherever you read them. Aethel now works with the files in your Drive, including Google Docs, Sheets, Slides and Forms, the same way she already works with your calendar and mail. Her written replies render properly in the web chat and on Telegram, and the voice reads only the words. This update needs one more Google permission, so every Google account signs in once more; until you do, the Google features for that account are paused.

What's new

Google Drive, Docs, Sheets, Slides and Forms

• Aethel works with the files in your Google Drive. Ask her to find a file, read a document or a spreadsheet, list what is in a folder, or read the questions and answers of a form. Paste a Google Docs or Drive link and she reads that file.
• She creates and edits Google files. A new Doc, Sheet, Slides deck or Form, made in the folder you name or the one the conversation is about, and edits to an existing one in place, so the link you already shared keeps working: replace text, add a row above the Total, bold a header, add a slide, add a question to a form, put a picture from your computer at the top of a document.
• Spreadsheets keep their formulas. When she reads a sheet she sees the formulas as well as the values, so a Total stays a formula after she adds a row.
• Move, copy, rename, restore, and share. Sharing a file, and moving one to the Trash, are things she asks about first. She never permanently deletes a file: a trashed file stays in Drive's Trash for 30 days. "Anyone with the link" sharing happens only when you say those words.
• Excel, Word and PowerPoint files stored in Drive are downloaded, changed through the Office app on your computer as in the last release, and uploaded back as a new version of the same file, with the previous version kept.
• One more Google permission, and one more sign-in. Google Drive access needs its own permission. Aethel tells you when an account needs to sign in again: in chat, on the Dashboard, in Settings, and in the setup wizard. Follow the link, leave every box ticked, and click Continue. Until then Calendar, Gmail, Tasks and Drive are paused for that account, so nothing runs half-connected. On an account that signs in again, Aethel checks it is the same account, and refuses a different one instead of quietly replacing your connection.

Replies that look right

• Web chat replies render as real formatting. Headings, dividers, tables, nested and numbered lists, quotes, checklists and strikethrough draw properly instead of showing raw #### and ---. It works on your home network with no internet connection, in every browser, light and dark.
• Telegram replies show the same things in a way Telegram can display. A divider becomes a thin line, a table becomes a neatly aligned block that fits a phone screen (or one short block per row when it is too wide), checklists show ☐ and ☑, and a link inside a table stays clickable. A long reply with a big table is split so Telegram accepts it instead of refusing the message.
• Voice replies say only the words. Code, table pipes, footnote markers and formatting marks are never read out. A file path is spoken in full instead of being cut to its extension, a formula is spoken in words, and "less than" is said as words rather than a symbol.
• No maths notation. Aethel writes a calculation in plain numbers and words rather than formula markup.

Web chat

• A picture from the internet in a reply does not load by itself. It shows as a button naming the site; click it to load the picture. Pictures Aethel made herself still show at once.
• A placeholder like <Your Name> at the start of a line no longer flattens the rest of the paragraph; it shows as text and the formatting around it draws.

Fixed

Google Drive

• A new file says where it went. When Aethel creates, uploads, moves or copies a file, the reply names the folder. A file made with no folder named says so and offers to move it.
• A folder that does not exist is reported as a missing folder, before anything is created, rather than as a missing file after a form was already made at the top level.
• Spreadsheet formula addresses are right for a read that starts somewhere other than cell A1.
• A download lands in Aethel's received-documents folder when you give a relative path, and the reply carries the full path.

Telegram and voice

• A small table with a link in it kept its link. It used to lose the web address on screen and in history.
• A Windows path such as C:\Users\you\Documents\report.docx is spoken whole. Backslashes had switched on the maths reader, which cut the path to "C:.docx".
• List<String> and "a < b" survive in a spoken reply; only real HTML tags are removed.

Behind the scenes

• Usage statistics are filed under the build that produced them, so an update followed by a same-day fix no longer merges two builds' counts.
• A shorter internal instruction file for the AI, with the same rules.

v2.17.1, September 25, 2026 (BETA_v2.17.1)

A release about voice replies that arrive faster and sound like one voice, and about error messages that say what actually went wrong.

What's new

Voice replies

• In the web chat, a voice reply starts playing while it is still being made. Microsoft's free voice service sometimes slows to a crawl, and a spoken reply could take a minute or more before you heard anything. The web chat now plays the reply as it arrives, and the written reply appears with it.
• Every voice reply is one recording. A slow reply used to be cut into pieces recorded separately and joined, so the voice changed tone at every join. It is now always one continuous recording.
• The player tells the truth. It shows when the reply is loading, playing, paused or waiting for more sound, and if the voice file is gone it says "Voice not available" instead of showing a player that does nothing.
• On Telegram, the progress message keeps counting while a voice note is being made, and /cancel works during that wait: it drops the voice note and sends the written reply.

Touchscreens

• The on-screen keyboard no longer pops up by itself. On a phone, a tablet or a car screen, the keyboard used to open after every reply and cover most of the conversation. It now opens only when you tap to type.

Fixed

AI engines

• Antigravity's "Verification Required" is explained. Google has started asking Antigravity users to approve their Google account. Aethel used to call this an error it did not recognise; she now says what it is and, when Google sends one, gives you the link to approve the account. You can also open a terminal, type agy and follow the link it shows. It does not pause your scheduled tasks.
• Codex's usage limit is recognised again. A newer Codex writes its limit message with a curly apostrophe, which hid it from Aethel, so a plan limit showed up as an unknown error. It is recognised again, together with the time the limit lifts.
• A scheduled task that hits a limit in its last minute runs a minute later, instead of waiting six hours.
• Error messages show the real error. When an engine crashed with a long technical trace, the message could show only the trace and lose the actual error line. It keeps the error line now, and an error message never includes pieces of the job's own work, such as text from an email it was reading or its half-written answer.

v2.17.0, September 23, 2026 (BETA_v2.17.0)

A release about editing your Office files without wrecking them, and about telling you the truth when something outside Aethel gets in the way. Aethel now changes an Excel, Word or PowerPoint file through the Office app you already have, then checks the result against the original before she calls it done. When a browser task drops because another Chrome extension took over the page, she names that extension and tells you what to do. On a Mac, she can see which permissions are actually granted and tells you when a lost one is working again.

What's new

Office files

• Aethel edits Excel, Word and PowerPoint files through the app you already have. Asked to add a column to a price list, replace text in a contract, or duplicate a slide, she used to rebuild the file with a Python library, which silently drops things it does not understand: a logo in the sheet header, column widths, the print area, the currency format. One customer's formatted price list came back damaged that way while Aethel reported success, because she only checked the cell values. She now drives your installed Excel, Word or PowerPoint to make the change, on Windows and on Mac, and always saves to a new file next to the original (<name> (Aethel).xlsx). The original is never touched.
• Every edit is checked against the original before she calls it done. A separate check compares the new file with the old one, looking at header pictures, column widths and styles, merged cells, number formats, validation rules, the print area and page setup, and reports anything that was lost. When Office is not installed, she falls back to the old method but still runs the check and tells you what it found, so a loss is never silent again.
• Your open documents are safe. If you open something in Excel, Word or PowerPoint while she is working, she never closes or hides that app and never quits it with your document inside. A job that hangs is stopped without touching what you have open.
• Exporting to PDF works the same way, through the app, so the PDF looks like the file does on screen.

Browser

• When another Chrome extension breaks the connection, Aethel names it. Some extensions, most often a password manager's autofill, draw their own frame into a login page, and Chrome then cuts every other extension off from that tab. Until now Aethel could only say the page kept dropping. She now identifies which extension did it, tells you the steps for that extension (usually: turn off its autofill for that one site), and retries the page once you say you have done it. A problem report now lists your installed Chrome extensions so we can see the same thing.
• A click that died because of that gets one answer, not two contradictory ones.

Mac permissions

• Aethel knows which permissions she really has. Full Disk Access is now checked by a real protected read rather than by asking macOS, which sometimes answers wrong. In chat she can tell you what is granted, what is not, and the exact System Settings pane to fix it, instead of guessing at terminal commands or suggesting a reinstall.
• You hear when a lost permission is back. One customer re-granted Full Disk Access after an update and it worked, but the macOS 27 privacy list did not show Aethel and nothing told him, so he assumed it had failed. Aethel now sends one message when a data source that was refused reads again, the same way she told you it was refused.
• The startup permission check waits for setup to finish, so a fresh install hears about a missing grant on its first run rather than after the next relaunch.

Reporting a problem

• "Report this to the developer" files a problem report. Asked to report something to the team, Aethel once found a contact with a similar name in the user's Gmail and emailed them. The team is reached only through the built-in problem report, and a message to a first name that matches more than one contact is confirmed with you before it goes.
• A report's crash log is capped at its newest few megabytes, so a laptop with old crash dumps no longer uploads 26 MB per report.

Fixed

Office files

• A column map for one sheet no longer shifts the others, and a deleted column is reported as removed rather than compared with its neighbour.
• A file saved as .xlsx from a macro-enabled original is reported as having lost its macros, not as all clear.
• A save that fails part way removes its half-written file and says so.
• On a Mac, a hang after the save keeps the finished file, and a timeout at the very first step names the macOS Automation prompt that is waiting for you.

Browser

• Every extension-caused drop is counted, not only the first one in a session, and a drop on one site is never blamed on the next.

v2.16.1, September 22, 2026 (BETA_v2.16.1)

A release for Windows computers whose Microsoft runtime is too old, and for finding out why an AI engine failed when Aethel cannot tell on its own.

What's new

Windows

• Aethel installs the Microsoft component it needs by itself. Voice transcription and memory search by meaning are built on the Microsoft Visual C++ runtime. Until now Aethel only checked that some version was present, so a 2016 edition passed the check while the search engine could not load. Aethel now checks the version. The update installs the current runtime, and if it is missing or too old later, Aethel installs it at startup. Windows asks for permission once; if you click No, Aethel waits a week before asking again.
• The installer offers a restart when Windows needs one to finish installing that component.
• A problem report names the exact file Windows could not find when the search engine still cannot load, instead of a generic message.
• Every screen says "out of date" rather than "missing" when that is the truth, and the Settings row that offers the install refreshes with the status above it.

Diagnostics

• An unrecognised engine error tells us what it said. When an AI engine fails in a way Aethel does not recognise, the daily usage report carries a short sample of the engine program's own error message, with paths, addresses and tokens removed. It never includes your messages, files or contacts. The privacy policy describes it.

Fixed

Windows

• Running the installer by hand over a running Aethel no longer records a false crash.
• The day's automation catch-up is no longer cancelled when the setup wizard reopens for a missing component.
• A refused permission prompt is recognised in every Windows language, not only English.
• An install that cannot start (a blocked download, or offline at startup) tells you once and tries again next time; a real installer failure stops the prompt from returning at every start.

Memory search

• After the runtime is repaired the search engine restarts on its own. The "restart Windows to finish" message is sent once, and only when a restart is what the failure needs.

Setup

• The wizard's final save lands before Aethel's background service starts, so the first run never reads setup as unfinished.

v2.16.0, September 20, 2026 (BETA_v2.16.0)

A release about the things Aethel does when you are not looking. The nightly maintenance run now checks that your automations are still working and tells you when one has stopped, reads your calendar, task list and the week's mail every night before it decides whether there is anything worth suggesting, and tidies memory instead of only pruning it. On Windows, a hard reset or a clock that jumped no longer gets filed as an Aethel crash, and a computer whose clock came back behind the server no longer locks your license over a key the server had just confirmed. Listing your email drafts works for the first time.

What's new

Scheduled tasks

• Aethel tells you when an automation has quietly stopped. The nightly maintenance run now looks at every scheduled task and notices one that has never run since you made it, one that is long overdue, one that keeps hitting the same wall run after run, and one that has been switched off for weeks. It tells the task's owner, and it does this even for someone who has not chatted with Aethel in a while, which is exactly the person who most needs to hear that their morning briefing died three weeks ago. A night your computer slept through is a skipped night, not a failure, and is never held against the task.
• The task list says whether each automation still works, in the app, in chat and in the Telegram commands, rather than only when it last ran.
• A task made in the app or in first run setup no longer reads as "two days old and never ran" on the night it was created.

Memory and the nightly run

• The nightly run reads your calendar, your task list and the last week of mail every night, read or unread, before it decides whether there is anything worth mentioning to you. It used to search only unread mail and open the calendar only when it already had a hunch, so someone who reads everything on arrival got no suggestions at all, and one engine went two nights running without opening the calendar once.
• Memory gets tidied, not only pruned. The nightly run now finds a person or project that has two pages, a fact saved twice, and two memories that contradict each other. Duplicate pages are merged, and a possible contradiction is put in front of you in the Memory tab's Decisions view, which until now could never actually show one. Choosing "Dismiss" or "Keep both" on a pair now sticks; before, the same pair came back every night until you deleted one of the memories.
• Open loops get closed. Something you mentioned once as unfinished used to live in memory forever, because nothing ever marked it done. The nightly run now closes a loop that has clearly been resolved, and stops folding loops into other pages where they could never be found again.
• A routine remembers when it was last used, so "you have not used weekly-report in two months" is something Aethel can actually know rather than guess.
• The nightly run can review a watcher, not only count them: one that has not fired in months is worth asking about.
• Conversation summaries are made with the engine's normal model. They had been forced onto a cheaper tier, which twice named a model the account could not use and silently stopped memory updating from that chat.

Email and documents

• "Show me my drafts" works. It had never worked: the request Aethel made was one the mail service refuses, every time, on every account. The list now also says when there are more drafts than it showed.
• A table Aethel calls "editable" is one you can edit. Asked for an editable table in a PowerPoint slide, she would sometimes produce a picture of a table and describe it as editable, because it looked the same on the slide. "Editable" now means what is inside the file, and a real check on three engines went from one slide in three wrong to none.

Updates

• Aethel checks for a new version every hour instead of once a day, so a fix reaches you the same morning it ships. A check is one small fetch; the installer itself is downloaded once per new version. The daily usage summary described in the Privacy Policy is now sent once every 24 hours.

Fixed

Windows: restarts, resets and clocks

• A hard reset or power loss no longer counts as an Aethel crash. Windows only records a shutdown time when it shuts down cleanly, so after a forced reset Aethel was reading an old shutdown, deciding no reboot had happened, and filing its own disappearance as a crash. It now also watches the computer's own uptime counter, which restarts at zero on any boot, clean or not. One customer's laptop was resetting itself and every one of those resets was reported to us as an Aethel crash.
• A clock that jumped backwards no longer fakes a crash, a healthy tray, or a stuck restart loop. The same resets brought that computer's clock back five and a half hours, which put Aethel's last heartbeat in the future. That made a dead Aethel look alive to the self-repair watchdog for hours, and it made a computer that had genuinely restarted Aethel a few times look like it was stuck in a loop, so nothing would start it again. Anything stamped in the future is now treated as unknown, and the watchdog restarts Aethel when the lock says it is gone. The /status command reads it the same way.
• The "start with Windows" hint finally clears. The check that verifies Aethel's self-repair entry could never match on any install, because Windows leaves out a setting when it is at its default value, so a warning about running on battery stayed on screen permanently. The re-check on every start also no longer deletes the half of the hint that was still true.

License

• A computer whose clock came back behind the server no longer locks Aethel over a license the server had just confirmed. Once the server has said yes in this session, that answer wins over a suspicious clock. A lock that heals on its own now resumes Aethel within thirty seconds rather than at the next daily check, and a key entered while Aethel is running no longer pauses it.
• An offline lock no longer says "in 999 days".
• Two license notices that said the same thing stopped arriving as a pair, and a slow clock produces one warning per session instead of filling the log.

Memory

• The nightly hygiene scan looks at the memories saved most recently, which is where a new contradiction against an old fact always is. It had been reading the oldest ones.
• A merge that adds nothing no longer rewrites the page, so a repaired page does not jump back to the top of every prompt as if it were news.
• The suggestions Aethel makes from your browsing name the device and the page they came from, and a downloaded file is named as it is, instead of the run claiming details it did not have.

macOS

• Aethel reads the computer's boot time once instead of asking the system for it every thirty seconds.

v2.15.0, September 17, 2026 (BETA_v2.15.0)

The largest release so far, and most of it is about things that were quietly not working. Voice got fast: a spoken reply now arrives together with its audio instead of a minute behind it, and Aethel answers a voice note in about a second and a half rather than six. Browser control stopped pretending: a browser link the extension never joined is reported in thirty seconds instead of four minutes, and it no longer opens tab after tab trying. On Windows, two features that had been dead on some computers for months, search by meaning and voice transcription, are now detected, explained and repaired by the update itself.

What's new

Voice

• A spoken reply arrives with its audio. The text used to appear first and the voice note followed whenever it was ready, sometimes a minute later, which made a short exchange feel broken. Aethel now waits briefly for the audio and delivers both together; if the audio is going to be slower than that, the text goes out and Aethel tells you the voice reply is on its way. The speech service also throttles roughly one connection in ten down to real time, which had turned some replies into a three minute wait. Aethel now opens a second connection when the first one is crawling and, for a long reply, speaks it in pieces over separate connections. A reply that took seventy three seconds now takes four.
• Aethel answers a voice note in about a second and a half. She used to start the speech recogniser from scratch for every note, which cost three and a half seconds before a single word was transcribed, whatever the length of the note. It now stays loaded for ten minutes after you use it and unloads itself afterwards, so it is not holding memory all day.
• She stops reading symbols out loud. Web chat and the voice tools were sending the raw reply to the speech service, so a stray asterisk was spoken, a web address was spelled out piece by piece, and a table was read as a wall of words. Everything spoken now goes through one cleaner: a link is spoken as the site's name, "/mo" as "per month", "~$5" as "about $5", a sum stays a sum, and a section break gets a pause you can actually hear.

Web chat

• Tab completes a slash command, and the Up and Down arrows recall what you sent.
• The composer fits a phone. The hint text fits the box and the send button stays on screen.
• The page follows new text only while you are already at the bottom, so scrolling up to read something no longer yanks you back down.
• A voice note in the history shows the words, not the markers around them.

Browser and computer control

• A dead browser link is reported in thirty seconds, and it no longer opens tab after tab. When the Bridge extension has not joined, Aethel now asks the connection itself whether the browser is there instead of inferring it from two and a half minutes of failures, and the answer is one tab rather than a pile of them. If you re-enable the extension, she reconnects on your next request instead of staying dead until Aethel is restarted. The Dashboard also stops calling a dead link "Connected", and the message you get names what is actually on your screen.
• "Take a screenshot into my Dropbox folder" puts the file in the folder. It was being saved somewhere else entirely, and the reply named a page the screenshot had never come from.
• A sequence of clicks and keystrokes stops when the window changes under it. You can now tell Aethel which app a sequence belongs to, and if something steals the foreground partway through she stops, instead of typing the rest of it into whatever is now in front.
• Aethel can tell when a window is running as administrator. She cannot click into one, but she had been reporting those clicks as successful anyway. She now notices and says so. That check had never once worked on a 64 bit install.
• A tab that goes blank is reported as a lost tab, instead of being described as a page that is still loading.

Memory

• Saving a note under a name you have used before updates that note instead of creating a second one with the same name. Updating a note also stops silently losing its title, which is what made the same briefing list exist twice with only one of them ever being read. Notes already split this way are repaired on the next start.
• Keyword search looks for the words you typed. It had been searching for your whole sentence as a single phrase, so anything short of an exact match found nothing.
• Search by meaning says out loud when it cannot start, in chat as well as on screen, rather than sitting behind a green tick. It also recovers by itself once the missing piece is installed, with no restart.
• A long conversation still gets summarised. Past a certain size the summary could no longer be produced at all, so memory quietly stopped updating from that conversation and the backlog only grew. Summaries are now taken in batches, and if one does get stuck you are told.

Windows

• The Microsoft runtime that two of Aethel's features need is installed by the update itself. Search by meaning and voice transcription are built on native libraries that will not start without a Microsoft runtime component, and on a computer without it both were dead while every check on every screen read perfectly healthy. One customer ran two months like that. The installer now carries the official Microsoft package and installs it when your computer is missing it, the first run setup offers it, and there is a button under Settings, System.
• Aethel stops disappearing on a laptop that is unplugged, and on any computer after three days. The two background entries Aethel creates in order to start itself were taking Windows' own defaults: do not start on battery, stop when the power is pulled, and kill anything still running after 72 hours. One laptop lost Aethel the moment the charger came out, and a desktop lost it every third day. Both entries are now created from an explicit definition with none of those limits, and an existing install is repaired once on the next start. A cancelled or failed update no longer leaves you with nothing starting at sign in.
• Croatian, Czech, Polish and other accented text survives a copy and paste. Text Aethel pasted for you was arriving mangled, or silently stripped back to plain letters, because it was crossing a channel that cannot carry anything but the basic alphabet. Checking the result by pasting it back gave a perfect answer every time, which is why this survived several rounds of testing.
• A password saved in the vault comes back out as it went in, accents and all.

Engines and connected services

• Aethel's runs see only Aethel's own tools. Whatever you have connected to Claude, Codex or Antigravity yourself is hidden for the length of an Aethel call and is never deleted or altered, and Aethel tells the AI it is not available so it cannot claim otherwise. Before this, another tool's broken or hanging connection could cost you the call.
• A connected service that needs a sign-in now signs in when you add it. Adding one used to appear to work and then quietly fail on every use, because no sign-in can be completed in the middle of a request. Aethel now checks each one and shows it as signed out or waiting for a key rather than "may work", and setting one up again under the same name runs the sign-in properly, which is the repair anyone would try first.
• The engine name next to a reply is the engine that answered it. Switching engine in one place used to leave the other place on the old engine for the rest of the day, while the label claimed otherwise.
• A task whose engine has run out of plan quota waits instead of failing every half hour. One install failed 140 times in a week, every one of them the same "your quota resets in 63 hours", which cannot change between two attempts thirty minutes apart. A task now pauses until something actually changes: you sign in, you change engine or model, the stated reset time passes, or six hours go by. You are told once when it pauses and once when it is working again, and the task list shows it as paused rather than healthy.
• Aethel tells you when the AI engine she is set to use is missing from your computer, and gives you the command to install it, instead of failing in the background where nobody sees it.
• A brief moment while an engine updates itself no longer costs you a reply, or stops your automations for six hours.

Email and calendar

• A reply quotes the original message, the way every mail program does it.
• Reading a whole conversation works. Asking Aethel to read an email thread failed about three times in five with "that email was not found". When a message genuinely cannot be found she now says which mailbox she looked in, and a scheduled task keeps using the account a result came from instead of drifting to another one.
• Aethel's optional activity notes cover every calendar you keep, not only invitations on your main one, and your own appointments count as well. On one real calendar with 72 events across six calendars, the old version saw two of them and reported that nothing was going on. A repeating appointment now counts as every time it happens rather than once.

Weather

• The forecast answers as many days as you ask for, up to sixteen, instead of always answering about today.

Fixed

Answering in the wrong language

• A foreign language email no longer changes the language Aethel talks to you in. The rule she had only covered your own messages, so a Croatian inbox, or a Spanish document she had just read, would carry over into her next reply to you. Writing something in another language on your behalf, such as an email to family, no longer flips the conversation either. The rule now travels with every request, including scheduled tasks, rather than only when your own words happen to be in front of her.

Scheduled tasks

• A task that could not finish quotes its own explanation instead of "it didn't say what stopped it", and a task that keeps hitting the same wall stops sending you the same message over and over.
• Deleting or renaming a paused task takes the pause with it, so rebuilding a task under the same name does not inherit the old one's pause.
• A task that actually succeeded is no longer recorded as failed because of an unrelated hiccup while it was writing its notes.

Chat and replies

• No more raw markup where a filename should be. Some engines wrap a file reference in their own markup, which was being passed straight through to you.
• A passing note between Aethel and Codex no longer appears as a tool called "Error". The progress card also shows what the AI is actually planning rather than a row reading "Todo list", and when Codex gives up on a request Aethel says why instead of calling it an empty reply.
• Aethel checks a document she made by reading it back, instead of apologising that she cannot look at one.
• An error nobody recognises is now named. A model your engine does not offer, a sign-in that has expired, and a provider having a bad ten minutes all used to arrive as the same unexplained failure.

Remote access, reports and reliability

• Remote access that has failed says which failure it is, and repairs itself once rather than restarting in a loop. One customer's remote access restarted every minute for four days behind a green light on screen.
• A message sent through your remote address survives the connection warming up instead of coming back as "No response".
• A save that Windows briefly refuses no longer costs you the message. Windows occasionally blocks a file from being replaced for a fraction of a second, usually because a backup or antivirus tool is reading it, and that moment used to lose whatever was being written.
• Removing a user from Aethel no longer risks other people's chat history.
• The Browser line stays on the Dashboard when remote access is down, along with everything below it.
• Problem reports now include your processor model and, on Windows, whether the Microsoft runtime components are installed, because both took a round trip to ask for and neither appears in any log.

v2.14.1, August 22, 2026 (BETA_v2.14.1)

If you run Antigravity as your engine, this one matters. On a busy conversation it was throwing your actual question away before the AI ever saw it, with no error and nothing in the log, so you got a confident answer to something you had asked hours earlier. That is fixed. So is a group of related problems that made Aethel answer about the wrong topic even on the other engines, and two that were quietly filling her memory with rubbish.

What's new

• One link that always gets you the current Aethel. tryaethel.com/get now offers the current build for Windows and for both kinds of Mac, and it follows the release automatically, so a link you send someone today still works after the next update. The page also offers a one-line install command for each platform. Those are worth using: a browser marks anything it downloads as untrusted, which is what triggers the blue Windows screen and the macOS block, and the install command does not, so neither warning appears. Both commands check the file is genuinely ours before installing it.

Fixed

Antigravity losing your question

• Your message was being cut off before the AI read it. Antigravity accepts only so much text at once and silently discards the rest: no error, no warning, nothing written down anywhere. Aethel puts your question last in what she sends, so on a long conversation your question was the first thing to go, and the AI answered whatever older part of the conversation survived. One machine ran like that for most of a day. Asked "can you send me links to each of those?" straight after a restaurant recommendation, Aethel replied about unread email, twice, once while quoting the restaurant message back. The standing instructions Aethel sends with every message now travel in a separate slot instead of competing with your words for the same space, which leaves your question far clear of the limit.
• And if a message is ever cut anyway, that run now counts as failed. Aethel notices, tells you, and throws the answer away, rather than handing you something the AI assembled out of half a question. This matters because the limit is undocumented and Antigravity updates itself every few days, so the number can move without warning.

Answering about the wrong thing

• Aethel now knows when her own last message was one you had not answered yet. A suggestion about unread email arrived at 12:41. At 12:50 the user asked "what do I have?", and Aethel answered about family activities in Tulsa, a conversation that had ended four and three quarter hours earlier. The suggestion was there in her notes, but nothing marked it as the last thing said or as something she had raised unprompted, so a three-word reply landed against the older topic instead. She now sees that her last message was an unanswered prompt from her, what it said, and when it went out.
• What she remembers about the recent past now carries its age. The panel of recent actions and the panel of what she has been working on both listed everything without a single date, so two minutes ago, five hours ago and nine days ago looked exactly alike and whatever happened to be at the top read as the live conversation. Everything in both panels is now dated. Picking a conversation back up days later works the same as before.
• She was reading chat timestamps in the wrong time zone. The times themselves have been shown in your local time since June, but the note attached to them still said they were in universal time and told the AI to convert. For two months that meant a second, unnecessary shift on every time it quoted back to you.
• A very long conversation can no longer crowd out your question. Your recent messages had a limit on how many were kept but no limit on how large they could get, alone among everything Aethel sends. There is now a size limit under the message count. It only engages on an unusually heavy day, and when it does the oldest messages are dropped whole and Aethel says so, rather than half-quoting an old reply and treating it as complete.

Memory picking up the wrong things

• Your "recurring topics" were the hidden code inside your emails. With the email observers switched on, Aethel had been recording things like "Recurring subject keywords in your sent mail: html, meta, doctype, head, charset". Most email is formatted, and the first part of a formatted email is instructions to your mail program rather than words anyone wrote, so that is what got stored. Seventy of the stored observations were of that shape and it had already spoiled her topic grouping. She now reads the words out of an email and ignores the formatting, and existing installs are cleaned up automatically on the next start.
• Pages about your projects and the people you know stopped filling up with browsing statistics. With the browsing observer on, a weekly summary of which sites you visited was being folded into the page about whatever project shared a name, so a page meant to record what a project is drifted into a browsing report, rewritten every week and never settling. Passive observations are no longer treated as things you told her. Pages already spoiled this way are repaired in place on the next start, without being pushed back to the top of her attention.

Problem reports and downloads

• A file you send with a problem report is no longer readable by anyone who has the link. The log and settings copy attached to a report were sitting on the same public address as the app downloads, so the only thing protecting them was that the link is long and hard to guess. That protection cannot be revoked and gives us no way to tell when a link has been used. Report files now require our team's credentials, including every report link already sent, and screenshots attached to a report are now sent as protected links rather than as pictures anyone could load. Downloads and update checks are unaffected and still need no sign-in.
• An interrupted download now picks up where it stopped instead of starting the fifty megabyte installer over from the beginning, and a browser that already has the file no longer fetches all of it again to check.
• There is a daily limit on problem reports, and hitting it now says so. It used to be reported as though the reporting service could not be reached, which sent people off to check their internet over a limit that clears by itself.

v2.14.0, August 21, 2026 (BETA_v2.14.0)

A scheduled task that could not do the job used to look exactly like one that did. From this release every run says what it actually achieved, and a run that hit a wall tells you instead of failing quietly behind a green tick. Web search was also completely dead in every packaged Windows build, which is fixed, and you can now manage your automations by asking rather than opening the app.

What's new

• A task now tells you whether it actually got the job done. Until now a task counted as successful whenever Aethel answered, and "I could not do this, the site made me reset my password" is an answer. One real check ran every weekday for a month, was stopped at a password screen on all 23 of those runs, wrote a clear explanation every time, and showed a tick on every screen; it was found a month later by opening the output by hand. Each run now ends with a verdict: done, waiting on you, couldn't finish, or nothing to do. The task list, the result window, the Dashboard and the /tasks command show that verdict instead of a plain tick, and a run that could not finish and told nobody sends you a message with the reason. A run that had already asked you for something, a code or an approval, stays quiet, because a second message about the same thing is just noise. Runs you start yourself with the Run button report the same way but never message you, since you are watching the result.
• You can manage your automations by asking. Aethel would happily rewrite a task you had built by hand, but refused to pause, rename or delete it, on the grounds that she had not created it. As the owner you can now say "pause my morning briefing", "rename it to Daily Brief", or "delete the invoice routine", and it happens, for any task or routine, not only the ones she suggested. She confirms with you first, because there is no undo. The maintenance task that keeps her memory tidy stays locked, and a task still waiting for your approval cannot be switched on through the back door.
• She looks in your bookmarks instead of guessing a web address. The bookmark and history search added in the last release was going almost entirely unused: asked to look something up, she would guess a company's address or dig through email, and asked point blank to look in your bookmarks she wrote a script to read the file by hand. She now reaches for it first. That matters because the bookmark you saved is the actual sign-in page for your account, while a guessed address is usually the marketing homepage, and a guess that happens to load looks like it worked.

Fixed

Web search

• Search was returning nothing at all in every packaged Windows build. One customer got an empty answer to every question for two days, including the single word "test", and was told each time that the web had nothing on it. The part of the search library that knows how to reach a search engine was never included in the Windows package, so every search silently fell back to a backup route, and for anyone whose network could not reach that route there was nothing left. Mac builds were unaffected. The missing piece now ships, both build machines are pinned to the same packaging toolchain, and the build stops rather than shipping if that piece ever goes missing again.
• "I found nothing" and "I could not search" are no longer the same answer. An empty result is a claim about the world, and Aethel may only make it when a search engine actually answered. When no search engine can be reached she now says search is down, and specifically does not tell you your question has no answer. A search that genuinely has no hits still says so.
• A search whose wording happens to include "no results found" no longer throws away a full page of real results.

When your AI plan hits its limit

• "You have hit your session limit" is explained, not reported as a mystery. Claude's session cap and Antigravity's individual quota matched nothing Aethel recognised, so you were told she had received an error she did not understand, followed by the raw message, for the most ordinary state a paid plan reaches. Both are now explained in plain words, with the time the cap lifts when the message carries one, where to check your usage, and the offer to switch to one of the other engines in the meantime.

Automations

• Two tasks whose names differ only in punctuation can no longer overwrite each other's instructions. "Standup" and "Standup!" were being stored in the same place. Renaming into that clash wrote one task's instructions over another's, so a task you had never touched would quietly start doing someone else's job on its own schedule. That is now refused, with an explanation of which name it collided with. The same fix closed a route where merely suggesting a task, which needs nobody's approval, could destroy a live task's instructions before anyone had approved anything.
• A task whose instructions have gone missing refuses to run and tells you, rather than asking the AI to do a job that is not there and scoring the result a success.
• A task you deleted in chat stays deleted, even if the Settings window happened to be open at the time and saved itself afterwards.
• Deleting a routine takes its parked suggestions with it, so approving one later cannot bring back the routine you just deleted.
• A run that was skipped no longer shows as a successful one on the Dashboard.
• Renaming a task whose instructions were typed straight into the editor no longer breaks every run of it from then on.
• A rename or delete asked for while that task is actually running is now refused until it finishes, instead of leaving a second copy running beside the first.

v2.13.0, August 20, 2026 (BETA_v2.13.0)

Aethel can now search the bookmarks and history your own browser already keeps, so "that article I read last week" is a question she can actually answer. Problem reports carry real evidence and real detail instead of a summary. Three of the most confusing things she did are gone: sending you to approve a card that does not exist, blaming herself for somebody else's broken add-on, and drawing her pointer hundreds of pixels away from the thing she was working on.

What's new

• She can find things you saved or visited. Ask for "my insurance login", "what's in my Medical folder", "the site I was on yesterday", or "what was I looking up about mortgages", and Aethel searches the bookmarks and history Chrome already keeps on your computer. Your folders keep their structure, so she can tell you which folder a link came from and list a folder you name. This is read-only, it never changes or deletes anything, it runs only when you ask, and only for you, the owner of the computer. Because you asked a direct question, the results are not filtered, so a search that matches something medical or financial returns it. Bear in mind Chrome quietly stores recent history from your phone and your other computers in the same place, so an answer can include those, and each result says which device it came from.
• A problem report now carries the evidence. When you agree to send a report to the dev team, Aethel can attach what actually shows the problem: a screenshot of the broken window, the file that came out wrong, the export that failed. She names each file and asks before sending it. She also has to write down what actually broke, in technical terms, rather than a one-line summary of it, which is the difference between a report we can act on and one we have to come back and ask about.
• What will never be attached is not left to her judgement. Your settings file, sign-in tokens, keys, and Aethel's own databases are refused outright, whatever a file has been named or where it has been moved to, including files that merely look ordinary but hold a sign-in inside them. If she refuses a file you asked for, she tells you.

Fixed

Add-ons and approvals

• She stops sending you to a screen that is not there. Asked to connect a service, Aethel would say an approval card had opened in the chat and ask you to approve it. No card existed and none could: the assistant she runs on has no window open in front of anybody. She now knows that, and will never send you to a card, a button, a menu command in that assistant, or a restart of it. Her own approve and reject commands are real and still work, in Telegram and in the web chat alike.
• A new add-on is actually tested, not just greeted. An add-on used to be reported as installed successfully as soon as it said hello, so one that could not perform a single action installed clean and was discovered to be broken weeks later, in the middle of something that mattered. Aethel now asks it to do one harmless thing and tells you plainly if that fails. When nothing safe can be tested she says it was not checked, and why, rather than implying it works.
• A broken add-on is named as the broken thing. When somebody else's add-on failed, Aethel treated it as her own defect: she filed several reports about one foreign bug, told the customer to report it to us, and worked patiently through the add-on's other features, every one doomed by the same fault. She now knows an add-on is a third party's software, says so, offers to find a replacement or remove it, and gets on with the rest of the job. She also remembers which add-ons are known to be failing, so the next morning's briefing does not start over.

Working on your screen

• The gold pointer lands on the thing she is working on. It was drawn wherever she guessed the page would end up after scrolling, which was hundreds of pixels wrong whenever the page was already at the top or bottom, or the thing sat inside its own scrolling panel. Measured on real pages, five out of five cases were 335 to 395 pixels off; all five are now exact. The pointer also stays put when what she clicked disappears, instead of flying to the corner of the window.

Answers you can rely on

• "Nothing in your history" now means nothing in your history. If Aethel cannot read one of your browser profiles because Chrome is writing to it at that moment, she says so and suggests trying again, instead of quietly answering from the rest and leaving you to conclude the page never existed.
• She does not claim you searched something on this computer when it may have been typed on your phone, and a search you ran two years ago is no longer reported as this week's just because you reopened the results page.
• A bookmark folder you have not filled yet still shows up when you ask what folders you have.
• A report is never lost with its attachments. If the evidence is too large for a slow connection, the report itself is sent without it rather than everything being thrown away, and a retry can no longer file the same report twice.

v2.12.0, August 19, 2026 (BETA_v2.12.0)

The release where Aethel learned to use a computer properly. Screen control has been rebuilt from the ground up after months of clicks landing in the wrong place, and she can now drag, draw, work across several monitors, and ask the computer itself exactly where a button is instead of squinting at a picture of it. Twelve more reported problems are closed, her nightly nudges finally draw on more than your email, and every message she sends you unprompted now says what it is.

What's new

• Aethel clicks where she looks. Screen control has been rebuilt around a single idea: the picture she is shown is the surface she works on, and the translation to your actual screen happens once, behind the scenes. Before this, she was doing arithmetic on a picture that was not the same size or shape as your display, and on a scaled or high-resolution screen the click landed somewhere else entirely. In plain terms: the thing she is aiming at is now the thing she clicks. Everything that depends on driving an app for you gets more reliable as a result.
• She asks your computer where the controls are. Rather than guess a button's position from an image, Aethel can now ask Windows or macOS for a list of every control in the window she is looking at, each with its name and its exact position. It takes a fraction of a second and it is exact rather than approximate, which is the difference between a sign-in that works and one that clicks next to the box.
• Several monitors. She takes her picture of the monitor holding the window you are actually working in, is told which one that is and what else is attached, and can be asked to look at a different screen. On a single-monitor machine nothing changes.
• Drag, draw and press-and-hold. She can drag something across the screen, hold a button down, sketch a shape, and drag an item into place inside a web page. Until now she could only click and type.
• A closer look at one window. She can photograph a single window instead of the whole desktop, so small text and small controls come through clearly.
• When a web page fails, she reads why. Aethel can now look at the browser's own record of what the page asked for and what came back, instead of speculating. Fewer confident wrong explanations.
• Every message Aethel sends you unprompted now says what it is. A nudge, a reminder, an approval request, a news item, an alert from something she is watching for you: each arrives labelled, as a coloured bubble with a tag in the web chat and a heading line on Telegram. They also become part of your conversation, so you can reply to one and she knows exactly what you are replying to.
• Nudges finally draw on more than your email. With every data source switched on, the nightly review was still building essentially every nudge from Gmail. Four of the six sources were quietly contributing nothing at all, and the richest of them had no way to spend what it had collected. The calendar source had been running for two months and recording nothing; it now sees your unanswered invitations and which meetings you tend to skip. Your file activity now includes your Desktop, not just Downloads and Documents. Browsing brought in pages synced from your phone and your other computers as though you had visited them here, and it counted Aethel's own browsing as yours; both are now sorted out, and the pages you return to repeatedly are the part that gets used. Settings also tells you plainly when a source you have switched on has found nothing, instead of a dead source looking identical to a healthy one.

Fixed

Working on your screen

• A step that fails partway through a drag lets go of the mouse. If something went wrong mid-drag, the button stayed held down, and the next thing you touched could be selected or dragged without you asking.
• Aethel refuses to type into her own window. A keystroke aimed at another app that arrived while Aethel's own window had focus used to be typed into Aethel.
• A refused screen reading no longer looks like an empty one. When permission was missing, the answer looked identical to a window with nothing in it, so Aethel drew the wrong conclusion and told you the wrong thing.

Browser

• One frozen tab no longer blocks everything. A single page that stops responding used to make every browser action fail, no matter which tab it was aimed at, with no way out except restarting Aethel. She now notices after a few failed steps, repairs the connection herself, and while it is happening she is told to keep going rather than to report that browsing is broken.
• Typing into a rich text editor works. Some page designs wrap the box you type in, and Aethel used to hand back a dead end instead of finding her way in.
• A slow page that says "not yet" is no longer treated as a broken one.
• When a browser step fails, the log records the real reason instead of a dump of the page.

Messages

• Quotes, spoilers and strikethrough now show up properly on Telegram, and a price written as "~$40" is no longer mistaken for crossed-out text.
• A code block keeps its formatting when a long reply is split across messages, and an unfinished one no longer leaks its markup into the text.
• Nothing is lost when Aethel is stopped mid-conversation. Messages that arrived at the moment she was stopped or restarted used to be dropped; they are now handed back and picked up on the next run.
• She replies in the language you just wrote in. With language set to Multilingual, a long conversation in one language used to pull her replies back into it even after you had switched.

Memory and email

• Aethel no longer passes off old notes as current news. When she looks something up in her memory she now sees when it was saved, so a briefing built on a note from three weeks ago is presented as three weeks old rather than as today.
• An email keeps the layout it was designed with. A message composed in one of the designed templates could go out in a different one, or fall back to plain text, so a carefully laid out email arrived looking nothing like the preview.

Privacy

• A password Aethel types for you leaves no trace. It could previously appear in the progress line on screen, stay behind on your clipboard (destroying whatever you had copied there), or be written into the log that travels with a problem report. All three are closed, and if your clipboard cannot be put back the way it was, it is cleared rather than left holding the password.
• Switching a data source off really does forget everything. Turning a source off, or pressing "Forget all observations", used to leave behind a short list of recently visited pages that was not visible anywhere in the Memory view.

AI engines

• Antigravity ending a request abruptly is now explained in plain words instead of being reported as an unrecognized error with the raw text attached, and a request that had not done anything yet is simply tried again.
• Aethel's own abilities attach reliably on a slow or heavily protected machine. On some Windows machines they failed to load in time, leaving her unable to reach your calendar or send a message for the whole conversation. She now waits long enough and tries once more if they do not appear.
• Accented and non-English text survives being written to a document on Windows. Croatian, Spanish and similar characters could turn into question marks, and Aethel would report success without ever checking what had actually been written.

Mac

• Restarting your Mac is no longer reported as an Aethel crash. Being asked to quit as part of a normal restart, or after re-granting a permission, was recorded the same way as a real failure.
• Reading the controls in a window works on macOS 26, where the previous approach came back empty.

Everything else

• Cancelling an update no longer leaves Aethel half asleep for half an hour. Closing the installer window without going through with it used to park her background checks.
• A scheduled task's name is shown to you, not to the person it is writing to. A task that sent a message on your behalf could leak its internal name into what the recipient received.
• Something Aethel is watching for you reports back to the conversation that asked for it, with the whole exchange in view rather than a fragment.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.11.1, August 14, 2026 (BETA_v2.11.1)

The largest fix release so far. Every problem testers reported since the last release is closed here, sixteen in all, along with two full internal reviews and a deliberate hardening pass for patchy internet, prompted by a tester working from a phone hotspot while travelling. There are four new things as well: replying to a message now carries its context, Telegram messages are properly formatted at last, Aethel can take over the passwords your browser has saved, and each installation now sends us a small daily usage count.

What's new

• Reply to a message and Aethel knows what you are talking about. Until now, using your chat app's reply function did nothing: Aethel saw only your new words, with no idea which earlier message you were pointing at. Now the message you replied to comes along as quoted context, in Telegram and in the web chat. Useful for "translate this one", "what did you mean here", or picking one item out of a long list she sent you an hour ago. In the web chat the quote is shown above what you are typing, and you can cancel it with a tap.
• Telegram messages are properly formatted. Bold, italics, lists, headings and links now appear as real formatting instead of the raw asterisks, hashes and brackets that used to show through. Long messages still split cleanly, and a split never lands in the middle of a code block or a link.
• Move the passwords your browser has saved into Aethel's vault. Chrome does not allow an automated click to pick a saved password, so any sign-in that depends on Chrome's password chooser is one Aethel simply cannot finish, no matter how it is worded. She now walks you through exporting your saved passwords once and imports them into her own encrypted vault, after which she can sign you in herself. The exported file is removed once the import succeeds, and if anything goes wrong she tells you where it is rather than leaving it lying around silently.
• Daily usage statistics. Once a day each installation now sends us a short summary: a random installation ID, the account owner's email and name as already set up in Aethel, the version, operating system, chosen AI engine, which features are on, and counts of activity such as how many messages were exchanged. Counts only, never content: no chat history, prompts, calendar, email, memory, file names or locations. It tells us how many people are running Aethel, on what, and who to help when something breaks. Full detail is in section 2.3 of the privacy policy.

Fixed

Working on a weak connection

• A patchy connection no longer breaks things quietly. This release goes through every path that touches the internet. Calendar and email calls retry a momentary blip instead of failing the whole job, mail is never skipped over during an outage, nudges are no longer lost when a send fails, reminders are no longer stranded half-sent, and no message is ever delivered twice as a result of a retry.
• A scheduled email check no longer fails on a one-second network hiccup. A tester's morning routine failed for exactly this reason: the connection wobbled for an instant and the whole task gave up.
• When the AI engine has a problem, you are told. A failure partway through a reply used to be swallowed and the entire request quietly run a second time, which cost time and could repeat actions. The error is now shown, and only a genuinely harmless case is retried, with a visible note when it is.
• Stopping something is no longer reported as a failure. Pressing Stop reported "no reply from the AI" or an outright error, and a cancelled scheduled task could send you a failure message you did not need.

Browser

• A slow website no longer costs you your tab. On heavy pages such as LinkedIn Sales Navigator, Aethel decided the browser connection was broken and restarted it, which closed the page you were on. She now waits out a slow page and only reconnects after a sustained run of real failures, and if she does reconnect she tells you which page was lost.
• A browser connection that never came up repairs itself instead of staying dead until you restart Aethel.
• Aethel starts faster and stops claiming the browser component is missing. The check for it could take up to half a minute on a machine with active security software, and a slow answer was read as "not installed", which blocked browsing entirely. The check is now effectively instant, and a successful install is believed the first time.

Messages, images and files

• Images and files Aethel creates now open when you tap them. Depending on where the file was saved and how the link was written, a picture could arrive as a broken link or a plain path you could not open. All the forms are now handled, including while a reply is still streaming in.
• The reply box on a phone behaves. The quoted message is readable on the coloured bubble, the cancel button no longer sits under your thumb next to Send, and a long quote no longer squeezes the typing area off the screen.

Voice and memory

• A damaged speech model no longer breaks every voice note forever. One interrupted download used to leave voice notes permanently failing; Aethel now spots the damage and repairs it herself.
• Aethel's automatic memory summaries were silently not being written on one of the AI engines after that engine's own update changed how it reports its models. Nothing was lost, but nothing new was being remembered from conversations either.

Windows

• Typing into an app running as administrator. Windows silently discards input sent to an elevated app, and Aethel used to report success while nothing had been typed. She now detects it and tells you what to do instead of leaving you to work out why nothing happened.
• The daily "access denied" message on non-English Windows is gone. Aethel's start-at-login check reported a failure once a day in the system language.
• Uninstalling now removes everything it should, and the documentation says plainly what is kept and what is deleted.

Everything else

• A Telegram message sent after a long quiet period arrives on the first try. After hours of silence, the first message out (including web chat sign-in codes) could fail once before getting through. Aethel now opens a fresh connection whenever the old one has gone quiet for too long.
• The "update available" notice always reaches you. If the message failed to send, or you use only the web chat, the notice used to be marked as delivered anyway and never retried. It now retries until it actually arrives, wherever you read your messages.
• An AI engine updated in the background is recognized right away. Aethel used to keep reporting the old version until restarted, and two things running at the same moment could clash over one engine's settings file. Both are fixed.
• One stuck tool no longer freezes the rest. A single tool that hung used to block every other tool behind it.
• A quiet task is no longer mistaken for a stuck one. Some real work produces no visible sign for a long time; Aethel now distinguishes "working silently" from "stopped", and timeout messages say what actually happened.
• A completed reply is never run a second time. On one engine, a reply that had already finished could be re-run from the start, doubling the wait and repeating any actions taken.
• A missing engine command is reinstalled rather than skipped over on every update from then on.
• Licence checks are steadier. A temporary internet outage no longer risks a licence problem, and a licence that was rejected once is rechecked properly when you are back online.
• The web chat no longer crashes at start-up in a rare timing case when Aethel was launching and reloading its settings at the same moment.
• The setup wizard falls back to a second download source when the usual one is unreachable, instead of failing the step.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.10.0, August 1, 2026 (BETA_v2.10.0)

Two new features and a broad round of polish. Aethel can now pass on occasional product news, and she chooses the right moment to send each nudge herself. Behind the scenes, this release also carries the results of a full review of everything shipped since 2.8.0: over forty small corrections across nudges, sign-in, licensing, and logging.

What's new

• Aethel news. From time to time we have something worth telling you: a new capability, an important fix, advice worth acting on. Aethel now checks for announcements about once an hour and passes anything new to you as a short Telegram message headed "Aethel news", mirrored into the web chat. The check is anonymous: your Aethel fetches a public file and sends nothing about you or your machine, so we never learn who read what. Each item is delivered once, expired items are never delivered late, and you can turn the whole thing off under Settings, System, Software update and news.
• Nudges now arrive when it makes sense, not inside a fixed window. Aethel used to hold every nudge for a set delivery window. Now she picks the moment herself, considering your local time, the day of the week, and what else is already queued. A reminder that only matters before your morning commute can arrive before it, and anything that would land at three in the morning waits for a reasonable daytime hour. Each nudge also knows when it stops being useful: if your machine was asleep past that point, Aethel either lets it go or reconsiders it the next day instead of waking you with stale news.

Fixed

• Scheduled work now follows your clock when you travel. Aethel already knew your timezone in conversation, but scheduled work quietly kept running on the computer's home clock. If you flew to Europe, a 7 AM briefing could still fire on Chicago time. Everything scheduled, including nudges and quiet hours, now follows where you actually are, using the same signal the chat already had.
• A nudge without a stated send time no longer fires at 3 AM. The nightly housekeeping run could deliver such a nudge the moment it was created, in the middle of the night. It now always receives a daytime slot.
• A licence could be wrongly invalidated during an update. If the one-time licence format upgrade was interrupted, Aethel treated a perfectly good licence as bad. The upgrade now either completes or leaves things exactly as they were.
• Remote sign-in privacy, follow-up. The previous release stopped publishing your account list outside your home network. This release closes a subtler leftover: the sign-in screen's replies no longer reveal, even indirectly, whether a given name exists.
• Signing in from your phone sticks. The web chat keeps you signed in for 30 days as intended, and switching between wifi and mobile data no longer piles up security-certificate questions on every network change.
• The microphone and camera work from in-car and other minimal browsers. Some browsers never even showed the permission prompt; the web chat now asks in a way they all understand.
• Stopping a running script actually stops it. The Stop command now reaches scripts the AI is running, instead of leaving them working invisibly.
• Closing the laptop lid mid-task no longer risks the task being cut short after waking. A lid close could quietly re-arm the "this run looks stuck" timer; waking now resumes the run with its full remaining time.
• The log file behaves. Log rotation had stopped working, every line was being written twice, and repetitive technical chatter was flooding the file. All fixed, which makes problem reports smaller and clearer.
• Crashes are recorded again. The catch-all that writes "Aethel closed unexpectedly" details was not being installed in the packaged app, so some real crashes left no trace for problem reports.
• The Dashboard no longer draws the News and Actions cards on top of each other.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.9.1, July 30, 2026 (BETA_v2.9.1)

A same-day follow-up to 2.9.0, fixing the new remote sign-in screen. If you use remote access, please update: signing in from outside your home network does not work in 2.9.0.

Fixed

• Signing in from outside your home network works again. The new sign-in screen asked for your name but gave you no way to submit it: no button, and the Enter key did nothing. You can now press Enter or use the Continue button, and leaving the box empty tells you so instead of appearing to ignore you. This only affected signing in remotely on a device that was not already signed in.
• It now tells you which name to type, and accepts your first name. The box said "Your name" with nothing to say which one, and typing just your first name was refused if your account is saved with a full name. Your first name alone now works, as does your full name or your Telegram ID, and capitals never mattered. The screen says all of this under the box.
• Your name is remembered on each device, so you only type it the first time on a given phone or laptop. It is stored on that device only.
• "Check for updates" tells you when the download is finished. It showed "downloading" and then never changed, even after the update was ready to install moments later. It now follows the download through and tells you to use the Install button at the top of the Dashboard.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.9.0, July 30, 2026 (BETA_v2.9.0)

A reliability release, and an unusually revealing one. A tester reported seeing "connection error" during normal use. Chasing that one message turned up a security problem in remote access, a false "your license has expired" alarm that had been quietly pausing a paying customer's app, and a timing bug we introduced ourselves in the last release. All are fixed here.

Please update. This release closes a security issue that affects anyone who has turned remote access on.

Security

• Sign-in from outside your home network is properly protected. When remote access was on, two things were wrong. The list of people who use your Aethel could be read by anyone who knew your address, and repeated failed sign-in attempts from a stranger could lock you out of your own account for fifteen minutes at a time. Neither is possible now: the account list is never published outside your network, someone else's failed attempts can no longer affect you, and entering your correct password always signs you in. Remote sign-in now asks you to type your name rather than picking it from a list.

Fixed

• Aethel no longer tells you your license expired when it has not. On machines where security software slows things down, Aethel could fail to read its own hardware identity, conclude the license was invalid, and pause your scheduled tasks, Telegram and web chat. One tester received twelve of these false messages in a month. Aethel now reads that identity a way that cannot time out, still recognises licences saved the older way, and repairs them automatically with nothing for you to do. Recovering from a genuine licence problem also no longer risks using up one of your activations.
• Long jobs get their full time again. A timing bug introduced in the last release could cut off a healthy task after twenty minutes and tell you it looked stuck, when in fact it was working the whole time and should have had the full two hours.
• "Connection error" in the web chat. A single dropped request used to leave a permanent error with no way to tell whether your message had been sent. Aethel now retries quietly, never sends or saves your message twice when it does, and if it truly cannot get through it tells you plainly whether the connection dropped for a moment or the app is not running. An error coming back from the server is now reported as what it is, instead of looking like Aethel is switched off.
• The web chat shows when it is reconnecting and clears the notice when it recovers, instead of sitting there looking frozen.
• Your browser stops asking about the certificate every time you change networks. Once you have accepted it, it stays accepted, and moving between wifi and a phone hotspot no longer undoes that on every device.
• The web chat address is easy to find again. Settings shows the network address alongside the name, and updates it when you move networks.
• "Python is not installed" when it clearly was. On slower machines Aethel could refuse to run scripts, insisting no usable Python was present. It now waits longer, tells the difference between "not there" and "did not answer in time", and says what it tried when it genuinely cannot find one.
• Large scripts work on Windows. Scripts the AI wrote past a certain size failed with a confusing filename error, with no rule anyone could work out. That limit is gone.
• AI-written scripts can reach secure websites on Mac. They could fail on every https address with a certificate error while everything else worked normally.
• Scripts stop when you stop them. Pressing Stop, or closing Aethel, now ends a running script and anything it started, instead of leaving it working invisibly in the background.
• Far fewer false crash warnings. Two out of three "Aethel closed unexpectedly" warnings were normal shutdowns, which buried the real ones. Shutting down, restarting and installing an update are now recognised for what they are.
• Opening Aethel twice in quick succession no longer opens two windows.
• Faster replies after some restarts. A startup timing problem could leave Aethel in a slower mode for the rest of the session, adding seconds to every request. It now recovers on its own within half a minute.

What's new

• Connection problems now get recorded. When your browser has trouble reaching Aethel, the details are saved on your own computer and included if you later choose to send a problem report. This class of problem previously left no trace anywhere, which is why it took so long to explain. It never includes anything you typed, any file name, or any address you visited.
• Better problem reports. A report now carries enough to tell the difference between your computer shutting down, Aethel being closed by something else, and Aethel crashing.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.8.0, July 23, 2026 (BETA_v2.8.0)

A release about letting big jobs finish: Aethel no longer cuts off long-running AI work at an arbitrary half hour, and a batch of fixes for problems our testers actually hit this week, including one that made the AI go completely silent for Antigravity users.

What's new

• Long jobs can finally finish. Every AI request used to stop at a hard 30 minutes, even when it was working productively the whole time. Now a run keeps going as long as the AI is actively doing things (using its tools, producing output), with a generous safety cap of 2 hours per request. In plain terms: a big research task or a long browser job gets the time it needs.
• Stuck runs are caught much faster. The flip side of the same change: when a run genuinely hangs, Aethel notices within minutes that nothing is happening and stops it, instead of letting it silently block scheduled work for half an hour. The message you get tells you clearly whether the run went quiet or simply hit the time cap, in plain minutes and hours.
• Sleep-friendly. Closing your laptop mid-run no longer makes Aethel wrongly decide the run was stuck the moment the machine wakes up.

Fixed

• Antigravity users: the AI answers again. A recent update to the Antigravity engine added a permission prompt that cannot be answered when Aethel runs it in the background, so every reply that needed a tool came back empty. The AI looked completely mute even though everything else worked. Aethel now grants its own assistant the permissions it needs automatically, the same way it already does for the other engines.
• "No account matches" when the account clearly existed. Scheduled tasks that named your connected Google account explicitly could fail with "account unknown" even though the same account was listed as available in the very same error. Related spots had the same blind spot: the AI could claim no email was connected right after you edited account assignments in Settings, and the account checkboxes could all show unchecked. All of these now recognize your account consistently.
• Accented file and folder names. Working with folders whose names contain accented letters (for example Croatian or Spanish characters) is more reliable: the AI now handles those paths with tools that treat them correctly instead of ones that garble them.
• Safer account setup. Adding a second Google account, or editing who may use which account, now stops with a clear message if Aethel cannot verify the current account first, instead of quietly leaving users without access.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.7.0, July 18, 2026 (BETA_v2.7.0)

A focused follow-up to v2.6.0: we found out why some testers still were not receiving suggestions even with everything turned on, fixed the root cause, and added alarms so this class of silent failure can never hide again.

Fixed

• Suggestions actually happen now. The nightly review that looks for helpful patterns had an instruction that quietly told the AI to skip the suggestion step whenever it could not verify your settings, which was always. Nights looked successful, but the suggestion step never ran. The AI is now handed your suggestion settings in writing at the start of every run, the skip loophole is gone, and every run must state what it did ("suggestions: 2 queued" or the reason there were none). Existing installs pick this up on their next restart.
• Your daily limit is respected upward, too. If you allow several suggestions per day, the nightly review now actually knows that number and can use it. Before, it silently assumed one.
• No repeats. A topic you were already nudged about this week will not come back day after day just because the numbers changed slightly. Something genuinely urgent, like a deadline arriving, still gets through.

What's new

• Alarms for silent failures. Three layers now watch the suggestion pipeline: each nightly run's result openly states whether the suggestion step ran, the System workers view in Settings shows a "Nudge analysis" row with a warning if it has not run in a few days, and if suggestions are on with fresh patterns waiting but nothing has gone out for 3 days, Aethel sends you a heads-up message directly.
• Clearer error trails. When part of the nightly review fails, the failure is now visible to the next run (which re-covers the gap), listed in the run's result, and never silently swallowed.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.6.0, July 18, 2026 (BETA_v2.6.0)

A release about Aethel's proactive side: suggestions (nudges) that actually arrive, arrive on time, and draw on better information, plus a new window into the background helpers that power them and a large reliability pass across the whole app.

What's new

• More than one suggestion a day, when there is more to say. If you allow several nudges per day in Settings, Aethel can now use that allowance: each suggestion covers a different topic, and when it has fresher information about something it already mentioned, the newer wording replaces the older one instead of piling up.
• Suggestions no longer expire on a stopwatch. A suggestion now lives until it stops being true ("pay by Friday" retires itself Friday night) instead of quietly vanishing after a fixed two days. Anything left unsent for a month is cleaned up automatically.
• See Aethel's background helpers. Settings, System now includes a "System workers" view showing the always-on helpers (memory summarizer, pattern watchers, suggestion delivery, cleanup) and when each last did real work. Nothing to configure, just visibility into what Aethel does for you behind the scenes.
• Scheduled work survives a sleeping Mac. Macs briefly half-wake on their own many times an hour, and Aethel's overnight runs could start during one of those blips and die mid-way. Aethel now recognizes those moments, waits for a real wake, and catches up properly, and a run that fails no longer counts as done: it retries.
• Suggestions know your places by name. Location-based patterns now speak in place names ("the gym on Main St"), never map coordinates, and turning the location watcher on now sends you a short Telegram how-to for sharing your live location, which is what makes this feature work.
• Clearer naming. The overnight task formerly called "Nightly Review" is now "System Maintenance", which better matches what it does. Existing installs rename themselves automatically.

Fixed

• Suggestions arrive when they should. A batch of delivery fixes: the daily allowance now resets at your midnight instead of drifting later every day, a recurring topic can suggest again after its first delivery instead of going silent forever, an undeliverable message stops retrying endlessly, duplicates can no longer slip through, and "today" deadlines are read in your timezone.
• Pattern watching got sharper. Inbox patterns now add up across the day, so "people you hear from most" fills in the way it should; download patterns survive a briefly disconnected drive; Chrome watching covers every Chrome profile, not just the first; a sign-in hiccup now recovers in minutes instead of hours; and the Safari permission warning on Mac clears itself once you grant access.
• Browser politeness. Aethel no longer opens a surprise setup tab in Chrome every time it starts, never closes a tab belonging to another automation tool you might use, and the little action labels near its cursor now stay up long enough to read.
• Memory keeps its good name. Long-running installs would eventually show a false "memory history may have been tampered with" warning after routine cleanup; that is fixed before anyone could hit it. Routine cleanup also trims more of the database bloat from day-to-day pattern watching.
• Updates show their progress. The in-app update download now shows live progress and never looks frozen, even when the size is not known in advance.
• Stronger scrubbing of uploaded reports. When you send a problem report, secrets are now scrubbed from every corner of your settings, not just the usual places, and the scrubber no longer mangles ordinary text that merely looks like a code.
• Honest sign-in status. On Mac, the dashboard now verifies your Claude sign-in is actually valid instead of assuming, so "signed in" means signed in.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.5.0, July 16, 2026 (BETA_v2.5.0)

A release focused on the browser: a clearer and faster experience when Aethel works on a website for you, and fixes for two frustrations that could turn a quick browser task into a slow, stop-start one.

What's new

• A clearer, more polished cursor while Aethel browses. The on-screen cursor has been redesigned in Aethel's gold, and it now tells you what it is doing as it goes: a small label names each action ("Clicking Send", "Filling Email"), a soft highlight frames the button or field it is working on, and you get live feedback while it fills out forms and dropdowns, not just while it clicks. If you would rather not see any of it, there is a new switch to turn it off under Settings, Browser Automation, "Show Aethel's cursor and highlights".
• Faster browsing. Aethel no longer stops to wait on its own cursor between steps, so clicking, typing, and filling forms move along noticeably quicker. The cursor also now appears on top of pop-up dialog boxes instead of disappearing behind them.
• Aethel remembers the page it left open. Over a back-and-forth conversation, Aethel now keeps track of the browser tab it was working in, so picking a task back up a few messages later, or even the next day, is smoother.

Fixed

• Aethel stops losing the tab it is working in. During a longer browser session, for example sending a batch of messages on a website, Aethel could lose track of its own tab and repeatedly ask you to re-share it, stretching a few-minute job into a slow, stop-start slog. It now recovers on its own and keeps going, and it no longer mistakes a slow-loading page for a broken connection.
• No more surprise "please add your tab" requests. Aethel used to sometimes ask you to open a tab and add it to a group before it could do anything, even when it could simply open the page itself. It now goes straight to the page, and only asks you to share a tab in the one situation that actually needs it: when you want it to work on a page you already have open.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.4.1, July 14, 2026 (BETA_v2.4.1)

A fixes-only release, driven directly by your problem reports from the last few days. Every issue reported since v2.4.0 is addressed here.

Fixed

• The Antigravity engine works again. A recent update to Google's Antigravity changed how it accepts instructions, which broke every Aethel request on that engine with an "empty prompt" error. Aethel now speaks the new format, and still the old one if you're on an older Antigravity. Long-running tasks also get the full time they need instead of being cut off after five minutes.
• Accented letters no longer turn into question marks. Documents and spreadsheets containing Croatian, Spanish, and other accented characters (č, š, ž, ñ, ¿, é, and friends) now keep every letter intact. In plain terms: Aethel now always writes files in the universal text format instead of ever falling back to your computer's regional one, which couldn't hold letters from more than one language at a time.
• Scheduled tasks use your latest saved information. A daily briefing (or any scheduled task) now reads the lists and notes you've saved in Aethel's memory each time it runs, so updating a list actually updates the briefing. Before, tasks couldn't see your memory at all and quietly worked from an old copy.
• "Prepare an email for me" now lands in your Drafts. When you ask Aethel to draft an email for your review, it creates a real draft in your Gmail, which shows up in whatever mail program you use (Gmail, Outlook, eM Client, Apple Mail), instead of improvising a file you'd have to hunt for.
• Upgrading from an old version on Mac is clean. If you upgraded from an early Aethel, the old app could stay behind, keep running, and keep answering your chats while looking like the new version. The installer now removes the old app, and Aethel also checks for it at startup, so there's only ever one Aethel answering you.
• Memory search survives upgrades. The smarter "understands what you meant" memory search now installs itself automatically after an upgrade instead of staying silently off until setup is re-run.
• Aethel browses in your open Chrome window on Mac. On a Mac, browsing tasks used to open a brand-new Chrome window every time, even when Chrome was already open. Aethel now joins the Chrome window you already have, and only opens one when none exists, exactly like on Windows. This works even if your screen is locked at the time.
• Error messages show the real reason. When an AI engine fails, Aethel now surfaces the actual error instead of the harmless warm-up noise that used to appear in its place, and a brief network hiccup no longer makes your Google account look disconnected until restart. Mac users are also now shown the Mac install command (not the Windows one) when an engine needs installing.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.4.0, July 9, 2026 (BETA_v2.4.0)

A feature release: memory that works in any language, calendar invites with a Google Meet link, a visible cursor while Aethel browses for you, smarter and better-timed nudges, and a large batch of reliability fixes for scheduled tasks, the browser, and web chat.

What's new

• Memory now works in any language. Aethel's memory search understands what you meant, not just the exact words, and that now works whether you talk to it in English, Spanish, or anything else. Existing installs upgrade themselves in the background, so nothing to do on your end.
• Calendar invites can include a Google Meet link. Ask Aethel to set up a meeting with a video call and the invite arrives with the Meet link attached, no follow-up in Google Calendar needed. It can also add a link to an event you already created, or take one off.
• Watch Aethel work in your browser. A small gold cursor now shows exactly where Aethel is pointing and clicking while it handles a website for you, so you can follow along instead of guessing. Under the hood it also reads pages more reliably, so multi-step website tasks succeed more often.
• Smarter nudges, at reasonable hours. If you've turned on nudges and some of the optional pattern-watchers, Aethel's suggestions now draw on more useful signals: bills and subscriptions it spots in your mail, messages sitting unread that look like they need a reply, calendar habits, and the kinds of sites you visit. Places are described by name, never as map coordinates. Nudges arrive at a natural, varied time of day, never at night, and you control how many per day. Requests for your approval also wait for daytime instead of pinging you at 3 in the morning.
• Pattern-watchers cover all your Google accounts. The optional email and calendar watchers now follow every Google account you've linked (or just the ones you pick), and there's a new optional watcher for incoming mail. Everything stays on your computer, as always. (See the Privacy Policy for the full detail.)
• Task changes take effect instantly. Adding, editing, or toggling a scheduled task now updates the running schedule on the spot, with no restart and no risk of a task misfiring right after a change.
• Clearer help when an AI engine is maxed out. If your engine hits its usage limit, Aethel now points you straight to that engine's own usage page so you can see where you stand.

Fixed

• Scheduled tasks run exactly once. After your computer wakes up or Aethel restarts, tasks no longer double-run or get falsely marked as skipped, and the nightly review catches up properly if the machine was asleep at its usual hour.
• The browser connection heals itself. If the link to Chrome gets stuck, Aethel now detects it quickly and recovers instead of sitting unusable. Dragging a file into a website's upload box now works, and clicks and screenshots are steadier, including on Windows screens with display scaling.
• Web chat survives a refresh. Reloading the page mid-answer no longer loses the reply in progress; it picks up right where it was.
• Smoother installs and updates on Windows. The installer now prepares Windows security settings before copying files, so installs and updates finish faster and without hiccups.
• No leftover pieces after a crash. If Aethel is force-closed, its helper processes now shut down with it instead of lingering in the background.
• Beta installs stay on the beta track. A beta build now always checks for beta updates, so beta testers get the next beta instead of waiting on a general release.
• One problem, one report. A hiccup could occasionally file the same problem report twice; it now files once.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.3.0, July 4, 2026 (BETA_v2.3.0)

A feature release: connect more than one Google account and choose who can use each, an always-current AI model list, a friendlier setup, and a big batch of reliability fixes for the browser, screenshots, mobile web chat, and messaging.

What's new

• Connect more than one Google account. Aethel can now link several Google accounts, for example a personal Gmail and a work one, or accounts for different family members. If you're the admin, you decide which of your authorized users is allowed to reach each account, and when more than one account could apply, Aethel asks which one to use so it never guesses. Setting up a single account works exactly as before. (See the User Guide for "Using more than one Google account.")
• Your AI model list stays current. When you pick a model for the backup AI engine, Aethel now reads the real, available list live instead of a fixed one, so the choices always match what your account can actually use.
• A friendlier setup. First-run setup now offers help right where you need it, checks things as you go (like confirming a connection code the moment you paste it), and tells you plainly why a step is blocked and how to fix it.
• Approve suggestions by number. When Aethel proposes a new scheduled task or routine and asks for your OK, you can approve or reject it by its number in chat instead of copying a long code.
• "What I've noticed." If you've turned on any of Aethel's optional pattern-watchers, the Memory tab now surfaces what it has noticed (like the people you email most, or where you tend to be), so you can see and manage it in one place.

Fixed

• The browser behaves. Aethel no longer leaves a trail of leftover tab groups piling up in Chrome, no longer opens duplicate tabs for the same page, and going "back" no longer hangs. When a website's upload box can't be used automatically, Aethel says so clearly and switches approaches instead of retrying in a loop.
• Screenshots are steadier on Windows. A screen capture that hit a momentary glitch used to fail outright; it now retries and recovers, so "look at my screen" is more reliable.
• Web chat looks right on your phone. The mobile layout no longer scrolls or zooms oddly, and a reply in progress no longer gets lost when it streams to your phone or a second tab.
• Clearer messages when an AI engine is busy. If the backup engine hits a rate limit or quota, you see the real reason instead of a blank "empty response."
• No more double messages. A hiccup while the AI was replying on Telegram could send you the same error twice; it now sends once.
• Stronger privacy in problem reports. When you choose to send a problem report, Aethel now also blanks out access tokens and email addresses from the uploaded log, on top of the passwords it already removed. As always this is opt-in and Aethel asks first. (See the Privacy Policy.)

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.2.0, July 1, 2026 (BETA_v2.2.0)

A feature release: a new Quick Actions tab, more you can do from chat, a heads-up if your license lapses, instant and accurate sign-in status, and a background service that looks after itself.

What's new

• One place for everything you reach for: the new Quick Actions tab. A single tab now gathers the common things that used to be scattered around the app: open web chat, open your Telegram bot, view the log, open your data folder, edit your settings, re-run setup, check for updates, restart the background service, and report a problem. The app is now called Aethel throughout, and the tray icon menu reads "Open Aethel" with a new "Report a Problem" option.
• More you can do from chat, without opening the app. From Telegram or web chat, /tasks now shows when each scheduled task last ran, /task <name> gives the full details of one task, and /routines and /routine <name> do the same for your saved routines. /engine is now a guided, numbered menu that walks you through switching your AI engine and picking a model, and if an engine you pick is not signed in or installed, Aethel tells you exactly where to fix it.
• You'll know if your license lapses. If your license expires while you are not in the app, Aethel now sends you a message so you find out and can renew, instead of quietly pausing with no explanation.
• Instant, accurate sign-in status. Aethel now checks whether each AI engine is signed in instantly and correctly, so the app no longer says you are signed in when you are not, or takes a while to tell you. The backup Antigravity engine is also more reliable.
• A background service that looks after itself. Aethel's always-on helper now tidies up leftover pieces from a previous run, notices if it stops unexpectedly, and can bring itself back, so it stays running more reliably in the background.

Fixed

• Setting up your first Telegram user just works. A catch-22 could stop the very first person from ever being authorized, so the message you sent to connect never went through. That is fixed. Telegram is now part of setup rather than optional, since it is how you reach Aethel and how web chat sends you your first login code.
• Buttons that open files now actually open them. On Windows, opening the log or the settings file could quietly do nothing when those file types had no default app. They now open in a text editor every time.
• Restart actually restarts. The "Restart background service" button now brings the service back even when it was stuck, and tells you honestly whether it came back rather than always sounding reassuring.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.1.1, June 24, 2026 (BETA_v2.1.1)

A patch release: a fix so Mac users actually get updates, smoother web file uploads, and problem reports the team can act on.

Fixed

• Mac updates work again. A naming mismatch meant Mac copies of Aethel quietly ignored every update. This is fixed, so Macs running 2.1.0 update normally from here on. If your Mac is stuck on 2.1.0, install 2.1.1 by hand once and updates resume on their own.
• No more wasted effort on web uploads. When a website asked Aethel to pick a file to upload, it could try over and over down a path the browser blocks, wasting time and money before giving up. Aethel now spots this right away and switches to clicking the file picker on screen like a person would, or asks you to pick the file.

Improved

• Problem reports the team can actually act on. When you ask Aethel to report a problem, it now sends the full log instead of a short snippet, plus a copy of your settings with all passwords and tokens blanked out. Aethel also keeps its log across restarts now, so a problem you hit before reopening isn't lost. As before this is opt-in, Aethel always asks first, and it never sends your chat history, files, or memory. See the updated Privacy Policy.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.1.0, June 21, 2026 (BETA_v2.1.0)

A feature release: Aethel can now keep itself up to date, web chat continues across your devices, and scheduled tasks can catch up on runs missed while your computer was off.

What's new

• Aethel updates itself. When a new version is available, Aethel now shows a notice in the app and can download and install it for you, then reopen on the new version. No more downloading and running the installer by hand. You can check for updates any time from Settings, and "Later" reminds you again tomorrow instead of hiding it for good.
• Web chat picks up where you left off. If Aethel is part way through a reply and you switch from your laptop to your phone, or open a second tab, the answer now continues live on the new device instead of leaving you waiting or starting over.
• Scheduled tasks can catch up on missed runs. If your computer was asleep or off when a task was due, Aethel can run it once the machine is back. It stays off by default so nothing runs unexpectedly; turn on "Catch up if missed" in a task's settings for the ones where a late run beats a skipped one.

Fixed

• No more false "Codex limit reached". When Codex hit a brief, temporary usage limit, Aethel could make it sound like your whole plan was used up even when plenty was left. It now explains it is a short, temporary limit, and when it clears, so you are not misled into thinking you are out of credit.
• Smoother backup AI engine. When you switch to the Antigravity engine, Aethel now finds it reliably even on unusual setups, and waits a little longer for its reply so it does not give up too early and come back empty.
• Faster first-time setup. Setting up the on-device search feature during onboarding could hang for several minutes. It now finishes in seconds.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.0.3, June 19, 2026 (BETA_v2.0.3)

A patch release: smoother browser automation, correct handling of "today" and "tomorrow", and steadier scheduling and updates.

Fixed

• Aethel knows what day it is, even days later. When picking up an older conversation, words like "today" and "tomorrow" could drift and point at the wrong day. Aethel now reads each past message against the day it was written and works out the real date, so reminders and plans land on the day you meant.
• Browser logins just work. When Aethel lands on a sign-in page that's already filled in with your saved details, it now simply submits it instead of stopping to ask. For an everyday task it goes straight to the page and gets on with it, and only asks you to share a tab as a fallback when it genuinely needs the one you already have open. If a site shows you as logged out, Aethel now points out you may not be signed in on this computer.
• Fewer leftover browser tabs. After the browser connection reconnects, Aethel could leave stray empty tabs behind. It now leaves fewer of them and explains the one it keeps for its connection.
• Browser automation stays connected. A recent change to the browser tools could make Aethel restart its browser connection mid-task. That's fixed, so multi-step web actions run through without interruption.
• New scheduled tasks don't fire on the spot. Creating a daily task in the afternoon for a morning time used to run it immediately, as if it had missed that morning. A task now only catches up on runs from before you created it, so a brand-new task waits for its next real time.
• Updates keep your setup choices. When you run the installer to update, all of its setup options stay selected as before, so an update doesn't quietly turn one off.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.0.2, June 18, 2026 (BETA_v2.0.2)

A patch release: a more reliable web search, smarter handling when there's no ready-made connector, steadier browser automation, and a built-in way to report problems to us.

What's new

• Report a problem to the Aethel team. If Aethel hits a genuine fault, or you're clearly stuck because it can't do something it should, it now offers to send us a report so we can fix it. It's always opt-in: Aethel asks first and only sends if you say yes. The report includes your version, system, AI engine, a short description, and a redacted snippet of the log (passwords and other people's details stripped out), and never your chat history, files, or memory. See the updated Privacy Policy.
• Aethel can build an integration when there's no ready-made one. If you ask for something the connector marketplace doesn't cover, Aethel can look up the service's own API, save any credentials to your vault, write a small program to use it, and save that as a routine you can re-run, after checking the plan with you first.

Fixed

• Web search is more reliable. Searches sometimes came back empty even when results existed. Aethel now uses a sturdier search engine, so it finds what's actually out there, and it reaches for a web search when the connector marketplace comes up empty instead of giving up.
• Browser automation keeps your place. When Aethel was filling a form or clicking through a site and the page reloaded or navigated, it could lose the tab and your in-progress work and start over on a blank one. It now stays on the page you're working on.

Known limitations

• Unchanged from earlier releases: the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.0.1, June 18, 2026 (BETA_v2.0.1)

A patch release: reliability fixes, a faster Memory section, and polish from live beta testing.

What's new

• Work with files anywhere on your computer. Aethel can now read, create, and edit files wherever they live on your machine, not just its own workspace, so you can point it at a document on another drive or a shared folder and it just works. It still asks before deleting or overwriting anything. See the updated Privacy Policy and Terms for what this means.
• Tidies up after itself. When Aethel sets up something temporary to get a job done (a preview page, scratch files), it now offers to clean it up once you're finished, so nothing is left running or cluttering your folders.
• Tells you when something breaks. If Aethel runs into a genuine glitch, it finishes what it can and then tells you plainly what went wrong, how serious it is, and that it's worth reporting, instead of failing quietly.
• The gold logo reaches your inbox. Emails Aethel sends now carry the gold AETHEL logo, matching the rest of the new look.
• See what your scheduled tasks did. Each scheduled task now shows when it last ran (and whether it succeeded) right in the Automations list, with a "Result" button to read back exactly what Aethel produced on its most recent run.

Fixed

• The credential vault is solid again. Saving a login now works reliably on Windows, and showing, copying, or saving a password no longer freezes the app. It could hang the first time you opened a saved login after launch, on both Windows and Mac, that's fixed.
• Browser actions are reliable again. Clicking and typing on web pages could fail on some sites after an automatic update to the browser tools. Aethel now works with the current tools and watches them at startup so it can warn early if they change again.
• Easier-to-read activity. When Aethel uses a tool, the small activity lines now show a file by name instead of a long folder path, so you can tell at a glance what it's doing.
• The Memory section is fast again. The Memories, Tool calls, and Decisions lists used to slow down as your history grew. They now open instantly and load more in small batches as you scroll.
• Stays current on its own. Aethel keeps its remote-access helper up to date automatically.

Known limitations

• Unchanged from v2.0.0 (below): the Mac build is not yet notarized and the Windows installer is not yet code-signed, Google sign-in still shows the "Google hasn't verified this app" screen, Aethel runs only while your computer is awake, and the Telegram bot has a 5-message-per-minute cap.

v2.0.0, June 12, 2026 (BETA_v2.0.0)

A major release. Aethel gained a real long-term memory, learned to run on-demand routines and act on what it watches, got a brand-new look, and got a full interface redesign, plus a lot of polish from a pre-release review and live install testing.

What's new

• A new look. Aethel wears a new gold AE mark everywhere: the app icon, the taskbar and Dock, the installer, the website. Icons are now drawn at every size they're shown at, so they stay sharp even tiny, and the tray icon adapts to your light or dark taskbar automatically, switching live when you change modes.
• One installer, three choices. Run the installer on a computer that already has Aethel and pick what you want: Install/Upgrade (the default, keeps all your data), Clean Install (erases everything and starts fresh), or Uninstall (removes Aethel and all its data). Same three choices on Mac and Windows.
• Living memory. Aethel now remembers across conversations: your preferences, the people and projects in your life (each gets its own self-updating page), and a running recap of where you left off, so you stop repeating yourself. A new Memory section lets you browse, search, pin, and forget anything she has learned, and trace where each fact came from. It all stays on your machine. A quiet background pass does the remembering on its own; just say "remember this" to save something on the spot, or "forget that" to remove it.
• Optional observers (off by default, at the bottom of Settings, AI, under Data Sources). If you turn them on, Aethel can learn patterns from your own activity: recurring email recipients, calendar habits, sites you visit often, download categories, location patterns. Privacy-hardened: filenames only (never contents), coarse location, and banking, health, and government sites are blocked outright.
• Routines. Save a set of instructions once ("weekly report", "file and log this") and run it on demand from chat or a task. It is the on-demand twin of a scheduled task, with its own view under Automations.
• Watchers that act. A watcher that spots a change can now do more than ping you: it can run one of your routines or follow a short instruction you set (for example, "summarize what changed and email it to me"), then report back.
• Remote access (off by default, under Settings, Web Chat). Optionally reach your web chat from anywhere at your own personal name.tryaethel.com address.
• Redesigned interface. A cleaner five-section sidebar (Dashboard, Automations, Memory, Settings, Logs) and a live Dashboard showing your connections, recent activity, and what is running at a glance.
• Custom instructions. Add your own standing instructions under Settings, Custom Instructions. Your words always win.
• Aethel knows your computer. She now knows whether she's on Windows or Mac, which version, and where your files live, so commands and paths come out right for your system instead of guessed.
• Easier permissions on Mac. Each macOS permission now has one Grant button: click it, approve the prompt, and the status updates by itself. Leftover permission entries from older versions are cleaned up automatically.
• The wizard keeps your AI engine fresh. The engine step now checks the AI apps it installed for updates and brings them up to date in the background.
• Quick commands, easy to find. Type / in any chat to see a menu of shortcuts. On Telegram it's the built-in command menu; in the web chat there's a new commands button, and the list filters as you type. New shortcuts let you list your routines, see what's waiting for your approval, look something up in memory, save a quick note, check your last shared location, and view your account details, all instantly without using up an AI request.
• Voice replies, your way. Tell Aethel /voice on to always reply out loud, /voice off for text only, or /voice auto to speak back only when you send a voice note.
• A better web chat on your phone. The web chat now lays out properly on a phone screen, the logo stays crisp and matches your light or dark theme, and you can start typing your next message while Aethel is still answering (it sends once she finishes).
• Smarter and more careful. Aethel remembers what she just did earlier in a conversation, plans before bigger changes instead of charging ahead, looks up your shared location instead of guessing (and never makes up an address), and offers to add a new integration when you ask for something she can't do yet. Her writing reads more naturally too.
• Nicer emails. Logos and images now appear inside the email body for every recipient instead of arriving as attachments, with cleaner formatting, sender names, and support for reply-to and message priority.
• Faster hands-on control. When Aethel is driving your browser or your computer, she now does more per step, so multi-step jobs finish quicker.

Changed

• Gemini is no longer an AI engine option. Google retired the Gemini CLI that Aethel used, so we removed it. Aethel now offers Codex (OpenAI), Antigravity (Google), and Claude (Anthropic). If you had Gemini selected, Aethel switches you to Antigravity (Google's successor, same Google sign-in) on the next launch. Your tasks, history, and settings are unchanged.
• Clearer names. "Tasks" are now Scheduled Tasks, "Plugins" are now Connectors, and reusable instruction sets are Routines. Note for testers coming from v1.0.0: this release changes a setting format with no automatic migration, so pick Clean Install in the installer instead of Install/Upgrade for this one update.
• Notes moved into Memory. The separate Notes page was folded into the Memory section. Ask Aethel to "save a note" and find it there.

Fixed

• A pre-release review went through every change since the first beta and fixed around thirty issues. Highlights: scheduled tasks could leave a duplicate "ghost" copy when renamed or deleted; a watcher whose action failed could go silent instead of alerting you; some settings changes did not take effect until a restart; an occasional weather lookup could drop the whole report; plus a set of Windows and Mac reliability fixes. For multi-user setups, a fix prevents a non-admin person's chat from briefly gaining elevated access while a background task is running.
• Live install testing caught and fixed several first-run issues: a fresh install now stays completely quiet until you finish the setup wizard (no early startup, no surprise sign-in windows), uninstalling from the installer now completes reliably, and the wizard's Authorize user step now works even when your Telegram account was already authorized by a previous install.
• When a data source is unreachable (like the weather service having an outage), Aethel now says so or finds the information another way instead of guessing, and she no longer wastes time retrying a service that's down.
• Faster replies when tools are involved: the AI now goes straight to the right tool instead of exploring its surroundings first.
• More polish from live beta testing: the web chat logo and the small mark next to replies now display correctly and switch with your light or dark theme, sending a voice note no longer occasionally returns the same answer twice, and /status now shows the model in use plus a single clean web address.
• On-device semantic search now repairs its own index if it was ever left incomplete, so memory search keeps working reliably after a fresh install.

Known limitations

• The Mac build is not yet notarized by Apple, and the Windows installer is not yet code-signed. First launch shows a Gatekeeper / SmartScreen prompt; see the Quick Start for the one-time approval steps.
• Google sign-in still shows the "Google hasn't verified this app" screen. Click Advanced then continue.
• On Intel Macs only, voice dependencies need Python 3.10 to 3.12.
• Aethel runs only while your computer is awake.
• The Telegram bot has a 5-message-per-minute cap.

v1.0.0 — May 1st 2026 (BETA_v1.0.0)

The first beta release. Aethel ships with everything you need for a personal AI operator on Mac and Windows: a guided setup wizard, three AI engines to choose from, Telegram and local web chat, scheduled tasks, browser automation, and a built-in credential vault.

What's new

• Setup wizard walks you through every piece on first launch: license activation, dependency check, Mac permissions, AI engine install + sign-in, Google sign-in, browser extension, Telegram bot, and your first scheduled task with a test run at the end.
• Three AI engines to pick from in Settings → AI: Gemini (Google), Codex (OpenAI), and Claude (Anthropic). Swap engines or models at any time without losing your tasks. (Gemini has since been removed, see the Unreleased entry above.)
• Telegram chat is the primary way to talk to Aethel from anywhere. Voice notes are transcribed and answered; you can send photos, documents, and your live location.
• Local web chat on your home Wi-Fi at https://<your-machine>.local:5004. Per-user passwords (claimed via a Telegram code on first sign-in), live streaming responses, voice in and out, file and image upload, light/dark theme, and multi-device sync — your phone and laptop see the same conversation update in real time.
• Tasks tab for scheduled jobs (daily / weekly / monthly / cron). Aethel can also propose new tasks based on patterns it notices; you reply /approve <id> or /reject <id> on Telegram to decide.
• Reminders tab for one-off Telegram pings and Watchers tab for inbox- and web-page-change alerts.
• Vault for your saved logins so Aethel can act on sites that need a password without juggling Chrome autofill. The encryption key lives in your OS keychain (macOS Keychain / Windows Credential Manager) — never on disk in plaintext.
• Plugins tab for installing extra MCP tools from the Smithery marketplace, and an Agent tab for longer-running multi-step jobs.
• Slash commands (/help, /status, /engine, /tasks, /reminders, /watchers, /cancel, /why) that bypass the AI when you just want a quick answer or to stop a long run.
• Progress card during AI runs that names the engine + model and lists the tools Aethel is calling, so you can see what's happening rather than waiting at a blank screen.
• Browser automation via the Playwright Bridge Chrome extension — Aethel drives your real browser with your active sessions, so multi-step flows (including 2FA) work end-to-end.

Known limitations

• The Mac build is not yet notarized by Apple, and the Windows installer is not yet code-signed. First launch shows a Gatekeeper / SmartScreen prompt — see the Quick Start for the one-time approval steps.
• Google OAuth still shows the "Google hasn't verified this app" screen during sign-in. Click Advanced → Go to Aethel (unsafe) to continue. Brand verification is in progress.
• On Intel Macs only, voice dependencies need Python 3.10 to 3.12 (newer versions break a few packages we rely on). Apple Silicon Macs and Windows have no such restriction.
• Aethel only runs while your computer is awake. Adjust your sleep settings if you want overnight scheduled tasks to fire.
• The Telegram bot has a 5-message-per-minute rate cap to play nicely with Telegram's API; bursting past that briefly queues replies.

Older releases will be listed here as they ship.