Clavus gets better
one release at a time.

After a Clavus Server update, every message and ritual round could fail at once with “MCP runtime cleanup could not confirm closure” until the assistant was restarted. The image zoom tool now starts from a path that stays the same across updates.

After a Clavus Server update, Codex sessions could fail to start because they still pointed at the previous version's files. Clavus now points them at the current version on its own.

When the assistant needs an API key or password, it now asks with a card: a hidden field, the sites the key may be sent to, and Save or Skip. The key goes straight into OpenClaw's protected store and never shows up in the conversation.

With OpenClaw 2026.9.7, Settings → Model access showed Connect even for providers you were signed in to, and Connect and Sign out failed. All three work again.

Typos that happen to start a longer word, like “teh” or “adress”, were taken for a word you were still typing and left alone. They are now corrected.

GPT-6.1 Sol replaces GPT-6 Sol as the main GPT model, at the same price. With a ChatGPT login it runs on your subscription (Clavus brings the newer Codex version it needs); with OpenRouter credits it runs through OpenRouter.

Simple questions now go to GPT-6 Luna instead of Gemma 4, and Luna is first in the model picker. Unlike Gemma it can use tools and see images, so rituals and image messages can run on it too.

If you share logs with a bug report, it now also says whether background tasks and rituals were running and includes recent server log lines, so problems on your server can be traced.

On a Mac trackpad, a single swipe to mark a conversation done could carry on into the next rows and mark several. One gesture now marks one conversation; a second swipe still works.

Right after the assistant restarted, a message sent while your model settings were still being applied could be dropped. Clavus now waits for the settings and sends once more if needed.

Editing or rerunning the first message of a conversation now picks the model and reasoning again, like a new conversation, unless you chose them while editing. Before, every later message stayed on Auto reasoning.

Instructions longer than 8,000 characters that the assistant passed to a background task were cut off and never ran. They now arrive complete.

A ritual round with nothing to report could start an extra conversation from the example in its own instructions, and you got a notice when it finished.

With OpenClaw's secret egress proxy turned on, the assistant got error 502 whenever it called Clavus itself, so a ritual could not list or create rituals. Local calls now skip the proxy.

After the assistant restarted, the model list could stay on the old state for up to 20 minutes. It now refreshes right away.

With a picture as wallpaper, every message logged an image error, and the picture could have been sent to the model along with your message.

When a transcription took longer than the Mac app was willing to wait, the dictation failed with “Transcription timed out” and the text was gone. Clavus now asks again with the recorded audio and picks up the transcription that was still running.

After some replies the main button kept switching between Stop and Done every two seconds. A reply that ends with tool calls and no text now counts as finished.

In any conversation, the assistant can now list, pause, reschedule and remove all scheduled jobs on your server, including ones that were not created as rituals. It used to send you to an administrator page that Clavus does not have. Anything new you schedule from Clavus becomes a ritual.

Select text on the keyboard or on your Mac and dictate what to do with it, for example “translate this into Spanish”. The selection is replaced with the result, also in as-is mode.

Start a dictation with the assistant’s name, like “Hey Jane, make this shorter”, and the text already in the field is rewritten, links included. Needs Clavus Mobile build 106 or Clavus Desktop 1.0.30.

Asking the assistant to change, merge, pause or retire rituals now works directly. A ritual can also be told to tidy up your other rituals.

Dictating “translate this” over a long selection replaced all of it with only the translated first part. Selections up to 20,000 characters are now handled whole.

A dictation spoken while another app was in front could end up in whatever conversation the Clavus window had open. It now goes to the conversation picker like any new message.

Turn on Open at Login in the menu bar menu or under Settings, Mac settings. If you switched it off in System Settings, turning it on takes you to Login Items there.

Start a dictation with the assistant’s name, like “Hey Jane, make this friendlier”, and Clavus rewrites the text already in the field you are typing in, links included.

Dictating “translate this” over a long selection pasted only the translated first part over the whole selection. Selections up to 20,000 characters are now sent whole.

The model list in the dictation widget now comes from your server, so it matches the app’s picker. Before, it still offered GPT 5.6 Sol and Opus 5.

Select text, tap the mic and say what to do with it, for example “translate this into Spanish” or “make it shorter”. The keyboard replaces the selection with the result. Undo brings the original back.

Start with the assistant’s name, like “Hey Jane, translate this into Spanish”, and the keyboard rewrites the text that is already in the field. A greeting to a person with the same name stays a normal message.

With text selected, the pencil button used to replace the selection with your spoken instruction. It now applies the instruction to the selected text. If the text changed in the meantime, your dictation stays in the keyboard instead of being typed into the field.

Images you attach now reach the model at up to 1568 pixels instead of about 600, and the original is kept so the assistant can zoom into it. On a screenshot of a dense table, answers went from 0 of 3 right to 3 of 3.

A message with an image now goes to the model you picked. Only Gemma, which cannot see images, hands it to your default model.

The document chip and menu in a conversation show the file extension, even when a long name is cut off.

Asked about your rituals, the assistant now checks the actual list of your rituals first.

On iPhone and iPad a PDF only showed its first page. Every page now renders and you can scroll through the whole document.

After the assistant changed a file, the viewer and downloads could keep showing the old version for a while. Files are now checked for changes every time they are opened.

On iPhone, tapping the edit chip could leave the screen stuck halfway between the conversation and the document. It now slides all the way, and the chip stays visible while the document is open.

A message sent with Send now on one device could be sent again from another device with the same conversation open.

A reply that was picked up again after reopening the conversation mid-answer could show its text twice.

With the Clavus keyboard on iPhone, a message you had just sent could come back into the empty composer.

The conversation info panel ran off the bottom of the screen when a conversation had many files. It now scrolls.

Link and file chips in replies no longer push the lines around them apart.

A title like "Pro & Contra" showed as "Pro & Contra" in the file list and on the document chip.

The keyboard notes when an app changes the text field on its own, like a chat clearing it after you send. This helps stop sent messages from coming back in more apps. It records only lengths and timing, never your text.

When you sent a chat message in Clavus while the keyboard was still correcting it, the corrected text could appear again in the empty message field. The keyboard now knows when Clavus clears the field and drops that correction.

A message sent with Send now while the assistant, Claude Code or Codex is working is now taken into the running reply after its current step, instead of stopping the run and starting over. DeepSeek sessions still stop and restart with your message.

The LLM Council now seats these three models, and Grok 4.6 and Gemini 3.8 Flash no longer take part. The model that calls the council leaves out its own seat, and models your install has no access to are skipped as before.

A message typed into an archived conversation that was open on screen, such as a bug report conversation, was dropped without a word. It now sends and brings the conversation back.

On installs where OpenClaw lives outside the standard system folders, Settings → Integrations and model sign-in failed with "spawn openclaw ENOENT". Clavus Server now also looks in the usual install locations, such as npm, Homebrew and nvm folders.

When you scroll through the suggested addresses in an email field, they now fade out at the edges instead of being cut off. At the start of the list the left edge stays sharp, and the right edge only fades while more suggestions are hidden.

In WhatsApp and other messengers, a correction that arrived after you had already sent the message could put the whole message back into the empty field. Once a message is sent, its correction is now dropped.

Report a bug now takes up to four screenshots, picked or pasted. In the conversation about a report you can send more, and the assistant looking into it can ask you for one.

Opening a conversation now lands at the start of the first message you have not read, even while it is still being written. With nothing new, it opens where you left off, and images that load late no longer push you somewhere in the middle.

A report now gets a short title describing the problem instead of the name of the chat it came from, so two reports from the same chat no longer look alike.

In a coding or support session, a message sent just as a run finished could show that run's whole reply a second time below it. The reply now stays above your message, shown once.

Typing a name with a space in the keyboard's emoji search, like "see no" for the see-no-evil monkey, now finds the emoji. Before, only single words matched and "see no" showed a swimmer and a wave.

On iOS 26 the emoji suggestions and search results above the keyboard looked frosted and cut off at the top. They now show clearly.

An emoji whose name matches what you typed now comes first. Searching "wink" used to show a trophy and a medal before the winking face.

Sharing a photo, voice memo or document into Clavus from another app now switches to Clavus with it attached in the composer. Before, the share sheet could stay open and the item only appeared the next time you opened Clavus.

Full-size photos from recent iPhone cameras no longer fail to share. Clavus shrinks them before adding them.

Clavus Server now installs a downloaded update on its own once nothing has been running for ten minutes: no reply being written, no task and no voice call. It also checks for updates every hour instead of every six hours. This update is the last one you install by hand.

A self-hosted install no longer waits for the OpenClaw Codex plugin update to use GPT-6 Sol. GPT-6 Sol, GPT-6 Astra, Opus 5.5 and Fable 5.1 are now the default models everywhere, each on your own ChatGPT or Claude subscription when you are signed in, and on OpenRouter otherwise if it is set up (with reasoning capped at high there).

Autocorrect, writing help and Clavus's own background helpers now use GPT-6 Luna instead of GPT-5.6 Luna. In our autocorrect test it fixed more typos and dialect spellings.

When Clavus updates in the background, the open app and the desktop overlay now wait until your voice conversation has ended before they reload.

When you ask something in a voice conversation that Clavus has to look up, it now reads out the answer instead of saying it could not retrieve it.

If Clavus cannot confirm whether a spoken request went through, it now says so and checks first instead of running it a second time.

When a report you filed is closed as not a bug, won't fix or duplicate, you now get the explanation and what to do next, not just a one-line verdict.

Long code in a message no longer shows two horizontal scroll bars on a Mac that always shows them.

When Clavus opens a page during a run, a live card under the action trail shows the tab as it happens, next to the chat on a computer and fullscreen on the phone. It folds to a single row once the run is done.

Settings now has an Integrations panel: add a service, sign in through its normal login page in your own browser, and reconnect or sign out from the same list. Sign-in no longer happens in a headless browser you cannot type into.

Rituals can now be scheduled for a single moment instead of a repeating rhythm, and Clavus can set one up for you from a conversation or from a running round.

The fullscreen button on a generated deck now works, including on iPhone, where the page bar and the composer step out of the way instead of sitting over the slides.

Opus 5.5 replaces Opus 5 wherever an Anthropic login is set up, with thinking always on. GPT-6 Sol replaces GPT 5.6 Sol on hosted instances now, and on a self-hosted install as soon as its OpenClaw Codex plugin reaches 0.155; until then that install keeps GPT 5.6 Sol.

Writing back to the conversation while a linked document has the screen to itself now works in the full-width view on a computer and on a phone held sideways, not just in the narrow phone layout.

The appearance step now closes by inviting you to name the colors or background you want, instead of leaving the prepared cards as the only choice.

After a gateway update, requests that need your approval expired instantly and the card never showed. Clavus now announces that it can handle approvals, so the card appears and waits for you.

When a slide deck is rewritten while you are reading it, the viewer stays on the slide you were on instead of jumping back to the first one.

Pressing Enter twice in the message box used to lose the empty line, show two carets and put the next characters on the wrong line. It now starts a clean new paragraph.

The highlighter needs a clearer scribble before it starts marking, so scrolling or a small wobble while reading leaves the text alone.

In a voice conversation, Clavus now waits for the finished answer instead of reading out an intermediate message that a newer one was about to replace.

When the diagnostic session behind a reported bug keeps failing, it is now given up after three attempts instead of relaunching forever.

Background tool runs now use their own scratch space and clear it out, so they stop filling the disk with leftover build files.

When the fast word check is sure something is wrong but none of the phone's suggestions fits, the language model can now suggest the intended word or split two words joined by accident.

After a fix is sent, the diagnostic channel stays available until the reporter verifies or closes the issue. Follow-up questions no longer fail just because the fix is awaiting confirmation.

When a ritual opens a new round, Clavus now hides the gateway's assistant-labeled copy of the prompt and shows only the actual round question. Reloading the conversation keeps it that way.

Tapping a task-list or table checkbox in a document no longer opens the on-screen keyboard when you were not already editing.

A word followed by a closing quote, bracket, comma, colon or semicolon is now checked like any other word once you type the next space or Return.

When the decision model is sure a word is a typo but none of the phone's suggestions fits, the server's language model may now name the intended word, or split two words that a stray letter joined, such as wholelimage into whole image.

Builds made with the newest Apple tooling could be closed by iOS right after opening. The app now uses the launch flow iOS expects.

A correction learned from repeated use now also counts when its target starts a sentence with a capital umlaut, such as Äpfel and äpfel.

A correction you have let stand twice for the same slip is applied on the phone from then on, without asking the server. Undo still wins and resets it, and a different correction starts the count over.

The keyboard's word check now uses Jev, a decision model from TypeSafe that answers with a probability instead of text, in about 0.3 seconds. It picks from the phone's own candidates or leaves the word alone, so dialect spellings, names and slang are no longer turned into dictionary words. Together with this, longer words accept a two-letter repair and suggestions from both dictionaries are considered.

Choosing whether a new message continues a recent conversation, and which model and thinking level it needs, now takes about 0.4 seconds instead of a second or more, using the same decision model. When it is unsure, the selector opens with only the plausible conversations.

Before a message is sent, the notes the workspace search found are checked once more for relevance, and only the ones that help are kept. Notes that share a topic word but say nothing about the question no longer take up space.

The provider had retired the Gemma 4 model, so simple requests silently fell back to a slower model. Gemma is now served through OpenRouter and quick replies are quick again.

Closing the main window with the red button or Cmd+W now hides it while Clavus keeps running in the menu bar. Opening it again from the Dock or the menu bar brings back your conversations instead of the first-launch setup page.

A word the keyboard once corrected can no longer become one of your trusted words by being typed often, even if a later check let it pass.

A slip of two letters in a longer word, such as a missing pair of letters in an eight-letter word, can now be repaired when the dictionary's best guess matches. Shorter words still only accept one-letter repairs.

Suggestions from the German and English dictionaries are now weighed together, so a German word is no longer crowded out by English guesses when you write German.

A word you have typed five times without it ever being corrected is left alone from then on. A word you have written three times can also serve as the repair for a slip next to it, in the spelling you use.

Links in a message waiting to send now appear as clickable links instead of raw Markdown. Lists and other message formatting render normally too.

Answers waiting behind an active reply remain paired with their questions in the queue and after sending, even when Clavus finishes the queue in the background.

Using Send now during a long reply no longer makes the running answer appear stopped or moves parts of it into an empty row. The reply stays live and its sections remain in order.

Signing in with OpenAI used to leave only GPT 5.6 Sol in the model picker. GPT-6 Astra now appears next to it with the same login, so Clavus can route thinking and writing tasks to it.

On installations with MCP servers configured, every new conversation failed within a second once 256 idle sessions had piled up, until the gateway was restarted. Idle sessions are now released after an hour.

Clavus can now send feedback to a finished Claude Code, Codex or DeepSeek session, run another round in the same context, and bring the updated result back.

Coding sessions now receive the Clavus guidance they can use, including your preferred tone, Vault files, rich replies and decision questions.

When only one model is available, Clavus now still follows your personalized reasoning guidance instead of silently choosing its own level.

Returning to a bug-report conversation now shows the latest reply, permission and case status instead of an old draft with an active Send button.

A table copied from a reply, a note, Notion or a web page used to land as one run-on line of cell text. It now shows as a table in the message field, scrolls sideways when it is wide, and is sent as a proper table.

The active Inbox, Open or Favorites chip is now a translucent film instead of a solid slab. With Reduce Transparency on it stays opaque.

The notice shown when diagnostic permission for a bug report ends now says plainly what happens next: nothing more goes out, what was shared stays with the maintainer, and a new report opens a new case.

Sources cited as path.md:14 showed a generic file icon and opened nothing because the line number was read as part of the file name.

A phone that went to the background right after sending could show the answer twice for about fifteen seconds when it came back.

When the connection dropped before the reply started, the automatic retry used to start a second, shorter answer and left the chat blank while Home already listed the real one. The retry now replays the original answer.

When Codex refuses every message of a session because it could not unload its previous configuration, Clavus resends once and, if that fails too, continues in a fresh session with the conversation restored.

Each conversation now gets 25 runs instead of 10 before asking for approval, and each approval adds another 25. Work paused at the old limit resumes automatically.

Reply links now offer Copy Link and stay linked when you select and paste them. The reply copy button pastes readable rich text and only confirms when the copy succeeds.

Open a conversation while its reply is still running and Clavus now reconnects to it, instead of looking idle and showing the answer only after the run ends.

You can now file a bug report from Settings without opening a conversation. The report explains, in the assistant's own words, what it shares and sends no chat content.

The weekly retro now measures many more features, including events, profile edits, and on-disk proofs, and only suggests a feature it can prove you have not tried. A feature that could not be measured never becomes a tip.

When the maintainer announces a fix or closes your bug report, the diagnostic permission you granted ends right away instead of running for the full seven days.

A typed or pasted address stays plain text you can edit like any word and becomes a link once sent. Labelled links keep their URL editor, which opens when you tap into the link and never takes focus while you type.

Segments of a running response stay in front of their browser row when a late final answer arrives, and a save that was overtaken by a delete can no longer bring removed rows back.

A sentence correction no longer rewrites across a web address or e-mail you typed earlier in the sentence. The correction starts after the link, so the link is never deleted and retyped. In the Clavus composer this used to wipe the rest of the sentence and drop you into the link's URL editor, with nothing to undo.

When the keyboard cannot verify a rewrite halfway through, it now types back what it had already removed instead of leaving that text gone.

Select any passage to comment, reply, resolve, or delete. Comment threads stay attached as the text moves, work on phones, accept dictation, and can mention another document with @.

The first table in a database document now gets the select, multi-select, checkbox, and stable ID controls declared by that document. Row edits are guarded so two writers do not silently overwrite each other.

Voice now uses GPT Live 1, keeps your profile and conversation order, and can hand longer work to your regular assistant without losing the result.

The weekly retro now reviews pending corrections, failures, wishes, models, and setup across your channels. It can also suggest one relevant Clavus feature you may not have tried.

Links in documents and comments now show compact destination chips. Recognized URLs get readable labels when pasted, and workspace documents open in place.

Saves from another tab or device now appear live and merge with your unsaved typing. Your caret stays in the same block or table cell while shared changes arrive.

Opening an edit from its conversation chip now scrolls the document to that exact change, while leaving the page still when the change is already visible.

New engine sessions start with Fable 5.1 or GPT-6 Astra at high effort. The LLM Council now seats Gemini 3.8 Flash, and optional DeepSeek choices include its latest lineup.

Bug-report case cards now show the reporter's name from their Clavus profile, so the maintainer can tell whose installation needs help.

After you stop a short response before it runs a tool, its partial text can remain visible, but that canceled turn no longer influences the next reply.

Starting Voice from desktop dictation no longer loses the handoff or lets a stale connection duplicate the finished text.

Replying inside a completed subtask now continues it as a normal conversation instead of treating the thread as closed.

The keyboard learns which email addresses, phone numbers, and other saved details you pick most often and suggests them first. Related address fields still stay together.

With the keyboard set to English, words like dont, cant, wont and im get their apostrophe right after the space, on the keyboard itself and without waiting for the end of the sentence. Words that are also correct without one, such as its, well or ill, are left alone. Undo keeps your spelling and the keyboard remembers that.

In editors that keep every line as its own field, such as the Notes page, the trash key and the field-wide rewrite could remove only the part of a long line the keyboard could see. They now act only when the whole line is readable and leave it untouched otherwise.

A word you finished while the previous one was still being checked was silently left alone, so quick typing left obvious mistakes standing. Every word is now checked on its own and corrected as soon as its answer is back, however far you have typed on.

On the Notes page the wand read nothing and changed nothing while the cursor wandered through the note. It now corrects the row you are in, and undo puts the text and the cursor back exactly where they were.

Swiping left deletes on the space bar again, exactly like on any other key: a short flick takes one character, a longer drag more. Moving the cursor is only ever a hold on space first, then a swipe; the multi-step undo brings back anything deleted by mistake.

Settings › iPhone settings gets a Keyboard haptics switch once the Clavus keyboard app supports it (build 78 or newer). The choice is stored with your profile, so it survives reinstalling the app.

A typed word is only replaced when the fix is a single-letter change to a spelling the keyboard already knows. Likely fragments from an accidental space are left for the sentence check, and applied, undone and redone corrections are recorded even while the connection is down.

Word fixes now accept Swiss ss spellings such as heisst when the dictionary offers the German form with ß.

On the installation that receives bug reports, typing in a case conversation now goes to the support session working on that case instead of the ordinary assistant. The reply shows as a Your call row and clears the waiting state.

A single-word autocorrection now applies even if you have already typed on. Before, the fix was dropped whenever the next keystroke came in before the answer, which at normal typing speed was most of the time.

The keyboard's undo button now steps back through what you did: typed words, deleted text, a cleared field, wand fixes and autocorrections, in order. Redo brings each step back.

Sliding left on the space bar now moves the cursor right away, like the system keyboard. Before, a slide that started before the trackpad hold could delete words instead, and a long text could be lost.

The 123 and ABC labels sit slightly higher on split keys for a more balanced appearance.

Upper characters on split keys sit slightly lower for a more visually centered appearance.

Brush over sentences or headings in a Markdown or HTML document, then send typed or recorded feedback against those exact passages. Review actions stay beside the document, with the original text available when you inspect or undo a change.

Clavus now asks assistants to speak plainly, avoid inflated phrasing and keep structure only where it helps. Rituals also pick up the same guidance.

Replies reconstructed from conversation history now keep commentary and actions in their original order and attached to the correct response.

A ritual that succeeds quietly after an automatic retry no longer leaves a failure message or asks for attention. The saved Home preview is repaired at the same time.

A one-time ritual now moves to Archive after its round is complete while keeping the result available in its conversation.

Requests from integrations that use the standard Responses message format now start their run at once instead of waiting for the conversation to be opened. Requests with images continue to use the path that preserves them.

Press Hyper+A with only a caret to improve the entire focused text field. Clavus restores your caret while it works and replaces the text only if the field is still unchanged.

On the hex keyboard, the special characters page now keeps the ABC key and return on the right, next to where the hex board has them, and drops the @ key from the bottom row. @ moved into the character grid. The traditional layout is unchanged.

On the hex keyboard, 123 now opens a number layer with the same honeycomb: digits in the middle, symbols on the sides, and punctuation, space, mic, and return where the letters keep them. ABC goes back to letters, and swiping up on it opens the special characters. The traditional layout and numeric fields keep the number pad.

Settings › iPhone settings has a new Keyboard haptics switch. Off silences every vibration the Clavus keyboard makes while typing, including long press and scrub gestures.

Early word corrections now accept Swiss ss spellings, such as heisst, when the keyboard dictionary suggests the corresponding German form with ß.

Early word checks use familiar spellings and a short list of nearby words. Likely fragments from accidental spaces stay unchanged for sentence correction.

Applied corrections, Undo and Redo are recorded even when the connection drops, helping identify unwanted changes without treating every untouched correction as approved.

Tapping the spelling-check spinner no longer starts a whole-field correction. Vocabulary downloads retry after connection failures and remove outdated imported spellings.

The keyboard can now check an unfamiliar word as you finish it, show that a check is running and apply a conservative correction before the sentence is complete. It skips name fields, remembers trusted spellings and treats Undo as your choice of the exact spelling.

The final special-characters page now includes & beside %.

Bug-report conversations now show an expandable list of the data sent, and diagnostic answers show their full redacted text. Masked personal details are counted by type before anything leaves your installation.

New reports start archived and return only when you need to answer, review a draft or see the outcome. Intermediate case updates stay out of the conversation.

Questions and diagnostic drafts now stay beside their controls, while messages and dictation from the composer go to the maintainer’s Clavus. Failed local checks send a safe reason and keep your draft.

Support case reviews now use a compatible Fable engine again. Repeated launch failures pause and ask for help instead of retrying every five minutes.

Clavus no longer shows a gateway outage when optional HTTP routes are disabled but chat is working. The check now uses the same authenticated connection as a real conversation.

Reopening a conversation while its reply is still running now reconnects behind the live work instead of leaving the new segments detached or out of order.

Cut and Delete work again after a floating note refreshes. Old editor listeners no longer swallow the shortcuts for the current note.

A successful ritual round with no reply now settles quietly instead of asking for attention. Malformed decisions still remain visible as failures.

Opening Insert or choosing an assistant in the dictation pill no longer brings Clavus to the foreground. Sending a result to Clavus still focuses it deliberately.

Clavus can now use GPT-6 Astra for complex reasoning and writing when your account has access. Automatic routing checks availability first, so it does not send work to a model you cannot use.

You now see the assistant’s live commentary in order while a task is running, including work that continues in the background. Updates stay attached to the correct run and no longer overwrite the browser’s copy of the conversation.

Clavus now gives every assistant the same guidance to lead with the result, use plain language and keep routine updates brief. Engine commands continue to work when that guidance is added.

You can now stop a ritual round or subtask directly from the message box. The red Stop button stays visible while work is running, so you can tell when the assistant is still busy.

In our English and German benchmark, MAI-Transcribe 2 returned transcripts 1.6× faster overall than Scribe v2, reaching 2.8× on the longest recording (74 seconds), measured from request to complete response. Scribe remains the automatic fallback; recordings up to the desktop limit of 20 minutes stay in one piece, and repeated closing sentences are filtered out.

Get Clavus now takes you straight to TestFlight on iPhone or the newest Mac app download.

The welcome introduction now uses the reply language in your profile, even when your browser is set to another language.

When a conversation exceeds the model’s context limit and cannot be shortened, Clavus now tries a fresh session with your thread history restored. Recovery is limited to once per thread every ten minutes, with a clear message if it cannot continue.

If your connection drops during an answer, Clavus reconnects to the work already running instead of sending your message again. It also detects a stalled connection sooner, preventing duplicate work, empty replies and stale error messages.

Turning your phone between landscape and portrait now keeps the conversation you were reading open. It previously sent you back to Home.

Questions now show their answer buttons even when the assistant accidentally formats the question as a code block. You no longer have to read the raw instructions in its place.

Support progress, diagnosis and plans now display proper formatting instead of raw asterisks and backticks. Remote images stay hidden so opening a report does not contact another site.

Notes no longer reports a merge with another device when Clavus has only added hidden metadata and the visible note has not changed.

A progress card shows the assistant’s plan, completed steps and current update while it works. It folds to one line when the answer is finished, and you can tap to review the steps.

Scheduled check-ins now appear in the assistant’s own words, and their replies start showing as soon as work begins. A conversation waiting for a check-in stays marked as running with a Stop button, so you can cancel it.

Interim updates now stay above the finished answer, so a completed conversation no longer looks as though the assistant is still working. The order holds when another device syncs an older copy.

Audio, images and other attachments now arrive with a player or file card you can open. Previously, you could receive only the caption.

Long answers no longer appear twice with a shortened copy underneath. Document edits near the end of those answers are now applied instead of silently skipped.

Reporting a conversation already titled “Bug: …” now gives the support conversation a clean title without adding a second bug-report prefix.

Each bug report now opens its own support conversation with questions, progress and fix updates, returning to your Inbox whenever there is news. Reports are encrypted so only the maintainer can read them.

With diagnostic permission enabled, Clavus can gather read-only diagnostic answers and mask sensitive details for your support case; anything that looks personal waits for you to send, edit or decline it. You can revoke permission at any time, and it expires automatically after seven days.

Using Hyper+A to improve selected text in Slack and other rich editors now keeps its links and carries its formatting along. Afterwards, your clipboard is restored, including any images and files.

When Clavus pastes an image for you, it now restores what you had copied beforehand, so it is still there for your next paste.

When the target field already contains a draft, dictation now preselects Insert so you can continue it. Choosing Insert also closes the menu immediately.

Dictating in the Clavus window with a document open beside it, you can now send what you have said so far and keep talking, with the passages you had selected travelling along. Until now the button and its shortcut only worked when the browser did the recording. This needs the Mac app at 1.0.24 or newer.

A Markdown file linked in a conversation now opens in the full editor even when it sits in a folder Clavus may not write to. It opens for reading only, and it counts as the conversation's document, so the assistant knows what you are looking at. Before, those files fell back to a plain preview.

The Fable slot now runs Claude Fable 5.1, in the picker, in routing and in the LLM Council. Adaptive thinking and the long context carry over.

The + card at the end of the Notes column opened its rows below the fold, under the composer on the Notes page and past the bottom edge in the panel. With a full column the row for creating a folder was off screen entirely. The card now brings its rows into view as it opens.

A message queued while the assistant was still writing appeared above the avatar, which read as if it had already been received. It now sits after the current turn, where it is the next thing to go out, and it scrolls into view when you are reading at the bottom.

When a run lost its closing signal, text the assistant had already written could stay out of the conversation until your next message brought it in. Clavus now waits for the run to finish and adds it there.

Models that comment between their tool calls had those comments sorted after the final text, so a finished answer ended on a line that read like work still in flight. The comments now stay in front of the answer, and the order holds through later edits to the conversation.

The model list in the dictation overlay names the Fable entry Fable 5.1, which is the model it now uses.

When you dictate in the Clavus window with a document open next to it, you can now send what you have said so far and keep talking. The Mac app hands over the section without a gap in the microphone, and the passages you had selected in the document travel with it.

After fifteen seconds on the spinner, a tap on the mic hands the keyboard back. Clavus keeps the recording as a retry entry, and if the transcript still arrives it fills that entry.

If iOS suspends Clavus while a dictation is still uploading, the recording is kept as a retry entry in the keyboard instead of being lost.

When Clavus was closed by iOS mid-transcription, the keyboard could keep showing the spinner with its mic disabled, and restarting the phone did not clear it. The keyboard now notices that nothing is being transcribed any more and hands the mic back.

If the agent Clavus was set to talk to had been renamed or removed on your OpenClaw install, every message failed inside the gateway and the app just kept spinning with no error. Clavus now checks which agents your install actually has and routes at the right one when the choice is unambiguous. When it is not, the conversation says so and names the agents you can pick from, instead of hanging.

In the browser, a chat with a document open beside it drew the wallpaper twice, mirrored at the divider. The window paints it once now, at any width.

In Markdown views a second, larger grey pill was painted behind the bar, and the document scrolled to a stop below the bar instead of passing behind it. The bar now floats over the page the way it does in a chat.

The one pixel border around the bar showed raw page content through it, so the bar read as a bare hairline against a busy background. The frosted film now reaches the edge.

A conversation now carries the document you last opened beside it in its bar. One tap reopens it, so you no longer have to find it again in your files after scrolling away.

Ask for a change to a document you have open and the text is rewritten in place, with the changed passages tinted so you can see what moved. Nothing waits for your approval. Tap a highlight to read what changed, keep it, or undo it.

During a long dictation next to a document you can cut the recording and send what you have said so far, instead of waiting for the whole session to end. Passages you selected in the document while speaking travel with the text, so the answer refers to the right place.

A round told to say nothing when there is nothing new could finish without any message at all, and the app then flagged it as needing you. A round with nothing to report now closes quietly.

A second copy of the app started for testing could adopt the real scheduled rounds and rewrite where they read their instructions. Instances now recognise each other by the gateway they talk to and stay out of each other's schedules.

On an OpenClaw install with several agents and ownership set to explicit, every conversation failed with a "no explicit owner" error before any reply arrived. Messages are now sent under the agent that owns the conversation, so these setups work again.

When a running agent asks you something directly, the question now appears as a card you can answer in the conversation. Before, the turn showed a bare tool step and looked like it had stalled while it waited for a reply that never had a place to go.

A thread with hundreds of tool steps kept every collapsed step in the page, which could grow large enough to crash the app on iPhone when the thread opened. Collapsed steps are now built only when you open them, and a long run shows the newest steps with a link to the earlier ones.

A message sent just as the previous run was shutting down could fail with a conflict error. It now takes over from the run that is on its way out.

Collapsing a reasoning block kept it closed only until the next update or a switch between answer versions reopened it. Your choice is now remembered for that message.

Starting a selection with the mouse inside a long answer opened the message actions and nudged the view. Clicks that land on text now just select it. Taps, long press and right click are unchanged.

When the gateway answered but rejected the request, the connection pill claimed the connection was lost, which pointed troubleshooting in the wrong direction. It now names the refusal and its reason.

With more than one agent configured, loading history and detecting a still-running turn silently failed, so sending looked fine while stalled turns never repaired themselves. Both now find the right conversation.

When OpenClaw updates itself to the 2.0 release, the gateway can fail to start because settings and stored conversations need migration. The server now notices this, migrates everything, and brings the gateway back on its own, usually within a few minutes and without any manual steps.

The app now works with both the current OpenClaw release and the new 2.0. Conversation history, answer recovery, and model settings carry over across the update.

When an agent asked a question with a text field, the keyboard or message box could cover the field you were typing in. The view now scrolls to keep it on screen.

Adding a line break at the bottom of a long draft could leave the cursor hidden below the edge of the message box.

After chatting next to an open document, returning home could leave you without a message box until a reload.

A ritual round that died used to end with a generic note. The round log and the round summary now carry the actual error, so you can see what went wrong without digging.

Routing, titles and other internal calls could inherit a faster, pricier delivery tier from the account default. They now always request the standard tier. Fast stays a per-message choice in the composer.

Conversations with very large histories no longer send multi-megabyte bodies to the phone during sync when the app would discard them anyway.

The iOS keyboard now receives your theme's accent color.

When the gateway accepted a message but never started the run, the app waited ten minutes and then asked you to resend. It now checks after two minutes and resends once by itself. A stop always stays final.

A run that completed while the connection was down could look unanswered and be sent again. Clavus now finds the finished reply and shows it instead of running the turn a second time.

If the server or the network path wedged before dispatch, the sending spinner could run indefinitely. The app now gives up after about a minute and hands the message to the normal recovery path.

When OpenAI has no free capacity for the chosen model, the turn ended with a confusing error about exhausted fallbacks. Clavus now says what happened, retries twice within seconds when nothing has run yet, and scheduled rounds switch to a fallback model.

After a session reset the assistant could lose the beginning of a conversation and claim it never happened. Missing turns are now detected on the next send and seeded back.

The animated avatar moved out of the title pill down to the end of the conversation, right above the composer, where you are already looking while you read. It shows whether she is listening, thinking or talking, and it keeps going through tool calls and streamed text until the answer is finished.

A message you could see on your desktop could be missing on your phone, with its answer still sitting there. Saving a conversation no longer removes anything the sending device happens not to know about, and your question is now stored the moment you send it, even if that tab closes right after.

After a restart, a running answer could stream into an old empty bubble further up the conversation while the turn you were watching stayed silent and its avatar slept. The reply now goes into the turn it belongs to.

A reply that quoted Clavus's own instructions could paint several attachment rows of sentence fragments under your question. An attachment is now only recognised when the line stands on its own and actually names a file or a link.

Rerunning the same turn twice forked a second branch of the same chat, and both ended up in the Inbox. Repeating the same rerun now opens the branch that already exists, while editing the message still starts a new one.

A favourite kept its star on the version the rerun replaced, so Home pinned the dead end right next to its successor. The star now moves to the branch that carries on.

The soft strip below the title pill sat on its border and blurred the line away, so the straight bottom edge was missing while the rounded ends stayed sharp.

Pressing a button on the dictation widget could pull the Clavus window in front of the app you were dictating into, so you had to switch back by hand. The widget now leaves the focus where it was.

The model picker in the dictation overlay still listed Opus 4.8, and it offered two thinking levels that the server does not answer for that model. It now shows Opus 5 with the levels that actually work.

If you have set a personal theme, the Clavus keyboard now uses its accent instead of the default clay: the wand and paste icons, a held letter and its option popup, the return key and the send button. It works out its own light and dark version of the colour, because the keyboard follows the app you are typing in, not the one Clavus itself is set to.

The return key and the selected option in a long-press popup drew their symbol in white. In dark mode the accent colour is light, so white on it was barely visible. The symbol colour is now worked out from the key colour and switches to dark ink where that reads better.

The desktop app keeps a short log of what it did with the microphone, and a bug report picks it up. Errors coming back from the app now say what actually went wrong.

A desktop dictation that ran past the twenty-minute limit was dropped without a word, and pressing Stop then failed. The recording is now finished and transcribed when it hits the limit, exactly as if you had stopped it yourself, and a transcription that fails can be retried.

Cards, lists, filter chips, ask cards and everything the assistant drops into a conversation now sit on a thin frosted film instead of a solid plate, so a picture wallpaper stays visible behind them. The film follows your theme.

Ritual threads show their round in the Home list, and the number no longer disappears when a long title runs out of room.

Pasted logs are no longer cut off after four thousand characters, and a report now includes what the app was holding right before a crash plus the desktop shell's recent log. Reports about crashes and dictation should be much easier to act on.

Opening a chat loaded its messages and then, milliseconds later, cleared them again: the title stayed up and the conversation went white. The chat you have on screen is now kept for as long as it is showing, whether you reached it from the list, a link or a second column.

When a stream died mid-answer, the reply had already arrived on the server but the chat kept an empty bubble, and asking again ran into a ten-minute timeout. Recovery now writes the finished answer into that empty bubble.

iOS can kill the app's web view and restart it in a loop, which showed up as a chat reloading every few seconds. Clavus now recognises that pattern at boot and stops the loop instead of riding it.

A model your install does not offer was shown as a broken red row in the connection test. It is now reported as off, and the test stops probing models that are not part of your inventory.

The schedule editor stayed open when you clicked next to it, because the frosted card around it trapped the dismiss layer.

The traditional keyboard now records the same anonymous touch positions the hex keyboard already does, so key targeting can be tuned for it later. Nothing you type is stored, only where keys were hit.

The space bar now reaches as far right as on the iPhone keyboard, so a right thumb no longer lands on the question mark. The return key is a little narrower to make room.

A conversation called New looks arrives with this update: pick your assistant's look and the app colors from a list with a preview next to every row. It brings two new designs made for you and four more to try (Moss Paper, Fjord Sky, Velvet Lines, Ink Paper). Tap a row and it applies right away, on every device.

When your assistant makes a picture, it appears as a row of thumbnails once the reply is finished. Tap one and the picture opens beside the conversation instead of covering it.

A set of illustrated wallpapers you can put behind the chat. You can pick one on its own without changing your colors, and any picture your assistant made can become the wallpaper.

Typing "zurich" now finds Zürich and "strasse" finds Straße. It works the other way round too, so the accented spelling finds the plain one.

The app loads one conversation at a time and asks only for what has changed since it last looked. Opening a long thread is quicker and the app keeps far less in memory.

During a voice conversation the avatar sits large at the bottom and moves while the assistant talks. Stop and Mute are together in the bar.

When your assistant offers you a set of options, they come as compact rows with a preview on the left and a tick on the one you pick.

Panels, menus, the overlay and the named accent colors follow your theme now. Notes take your colors too, both in the app and in the panel on its own.

A note row is no longer cut off at a fixed height, so you can read a long note without scrolling inside the row.

Auto-close stays off until you turn it on. It was on by default in the previous update.

Pasting text turns it into a file only past 10,000 characters, up from 2,000.

Expanding an embedded file opens it in the file pane, the same way a link does. The extra link that used to sit next to every embed is gone.

A note you were editing could be replaced by a copy fetched a moment earlier. Notes now ignore a read that is older than the version they already hold.

The assistant's scheduled wake-ups appeared as empty turns when a conversation was restored. They are hidden now, including in conversations that were already cached.

Sending a message while the previous one was still starting could leave the conversation waiting for nothing. A send that never gets going now expires and the queued message goes through.

It hung too low on some pages, and the strip beneath it read as a stray shadow band over a wallpaper. It now blurs into whatever is behind it.

The small dictation window takes the theme you picked in Clavus, instead of staying in the original gray.

The assistant overlay is built the first time you summon it rather than at launch, so the app is ready sooner and uses less memory until you need it.

The Dock icon stood taller than every other icon next to it. It now uses the same inset as the rest of macOS.

Pressing the Hyperkey while you dictate into the composer now presses Stop, so what you said is kept. It used to end the recording and throw the words away.

When you tint app icons on your iPhone, Clavus now uses a purpose-made monochrome icon and follows the exact color you choose.

A new conversation called Make it yours lets you rename your assistant, pick a look for it (a drawn face, a soft orb, a blob, or none) and choose colors: the Clay look you know plus two designs made for you. Tap a row and it applies right away, on every device. Ask in any conversation for warmer, calmer, or a color you like and your assistant adjusts it.

A small animated avatar sits next to the conversation title and marks the assistant's conversations on Home. It blinks, looks up while thinking, and reacts when something fails.

Describe tags in words ("Office is clients and offers, Private is family and trips") and Clavus files every conversation under them automatically. The tags appear as tabs next to Inbox, Open, and Favorites.

Tell your assistant to mark conversations as done a day after you read them, and it happens on all your devices. Favorites, unread answers, and running work stay open.

The browser icon and the icon a new home screen install gets now use your accent color. The Mac app's Dock icon follows too with Clavus Desktop 1.0.19.

Status lines and errors are written in the first person: "I'm thinking", "Something broke on my side". The raw error stays one tap away under Details.

When the assistant explains between two actions, that text now shows between the steps instead of below all of them.

Placeholders greet you by name and address the assistant by its name. Spelling across the app is American English now.

When you pick a color theme in Clavus, the Dock icon takes your accent color. Switching back to Clay restores the original icon.

When a task needs longer than one turn allows (a build, a crawl, a render), the assistant now starts it in the background and schedules a check-in. Clavus watches the process and reopens the conversation the moment it finishes. Before, the assistant waited inside the turn until the run was cut off after about 30 minutes and you got an Agent run failed instead of an answer.

A conversation with a pending check-in shows the process or timer it waits for and the job's latest progress line, with a button to discard the check-in and, for local processes, one to stop the job. The thread list marks such conversations as waiting.

Rounds were cut off after 30 minutes no matter how much work was left. The limit is now three hours.

On desktop, any linked file opens side by side with the conversation, not only Markdown notes. On the phone the document view has its own composer. Files and chat embeds refresh on their own when the file changes on disk.

Dialect words are no longer rewritten into Standard German by the keyboard's autocorrect.

The keyboard's calibration log now records when the personal layout adjusted a tap to a neighbouring key, so the adjustment can be checked against what you actually typed next and the layout keeps improving.

If the Clavus keyboard suddenly gives way to the standard keyboard, iOS has closed it for using too much memory. The keyboard now records its memory use while it is open and notes when it was closed this way, so the cause can be traced from the report instead of guessed. When memory runs high it also frees what it can before iOS steps in.

When the assistant asks you to pick from options that are not mutually exclusive, it now offers a multi-select instead of forcing a single choice.

A conversation forked from an earlier point keeps the model and reasoning level of the conversation it came from.

After the server updates, open tabs load the new version in the background and switch over without a blank screen, and never while you are typing.

A file you opened before shows right away from the saved copy and refreshes in the background.

Installing 1.0.27 on an existing server stopped with Install failed and left the previous version running. The update package is built differently now, and every release is checked for the problem before it ships.

Chips in notes no longer lose the gap between them or wrap in the wrong place.

An HTML file that fills the screen no longer shows a dissolve band at the top.

The spinner while voice mode connects uses the app's colors in light and dark mode instead of a fixed indigo.

The Clavus Conversation control in Control Center now behaves like the shortcut: the recording starts straight in the Dynamic Island, and Clavus opens only when you stop. The Action Button and the control stop each other's recordings. On iOS 18, the control needs Clavus to have been opened once since the last restart.

Tapping the Clavus Conversation, New Chat, or Rituals control did nothing. The controls now open Clavus and run the chosen action.

Tap a link in a note to see where it points, change it, or open it in place.

A bug report now includes the app's recent log lines, so problems that only show on your device can be diagnosed from the report alone.

A file open in the viewer refreshes by itself when something changes it, for example a running assistant session writing to it.

Reports and embedded pages draw edge to edge and scroll behind the top bar. Fullscreen embeds in chat use the same viewer as files.

Leaving a conversation you opened from a list page goes back to that page instead of home.

The composer's Photo button goes straight to the photo library, and Reference a file is now called Add a file.

A Claude Code or Codex session that finished its reply while background work is still running now shows Waiting instead of looking stuck. Stop keeps working the whole time.

If a run went quiet after doing real work, the error now says how many steps it completed, so you know the work may be partly done. Runs that never started say that, instead of a technical guess.

Links inside a card's text now become the card's link button instead of showing as raw markdown code.

If the app lost the live view of a running reply and you sent a follow-up, a second run could start on top of the first. The follow-up is now held back and the app reattaches to the running reply.

Turns covered by a subscription were sometimes counted as paid API cost in the conversation cost view.

Uploading a file with umlauts or accents no longer garbles its name.

The bar around the page title pill swallowed taps on things behind it.

Choosing Photo in the composer now goes straight to your photo library instead of first asking whether you want to take a photo or browse files. Camera and files keep their own entries in the same menu.

Talk to Clavus no longer opens the app or covers the screen while you speak. The recording runs in the Dynamic Island, and tapping its stop button opens Clavus so your words land in the matching conversation, a new one, or the conversation picker. If you use the shortcut through Back Tap or the Shortcuts app, re-add the Clavus Conversation action once so it picks up the new behavior.

Pressing the same Conversation button again while a dictation is running now stops it instead of doing nothing.

Starting a dictation while Clavus is already in front no longer shows the full-screen recording overlay. The composer's recording bar and the Dynamic Island carry the feedback instead.

The area at the bottom right of the screen that opens the Slidepad is now half as tall, so the panel opens less often by accident when the mouse passes along the right edge.

Every change to a file in your workspace is now kept, whether the assistant or you made it. Open a Markdown file or an HTML report, tap the history button in the header, and pick a version: the view switches to that version, with arrows to step back and forth and a Restore button if you want it back. Each version names the conversation that changed it.

Messages you queue behind a running reply used to wait until Clavus was open somewhere. The server now sends them itself, so a follow-up typed on the phone goes out even after you close the app.

Dictations from the iPhone quick entry are routed by the assistant to the conversation they belong to, with the home screen showing recent conversations to pick from.

Outside its cards and menus, the floating notes panel no longer blocks the window underneath. The Mac app side of this shipped in Clavus Desktop 1.0.17.

In notes, Tab and Shift+Tab indent and outdent the rows you have selected.

A document opened next to a conversation can be closed with Back in its own header. Before, you had to expand it to full width first.

A linked file that was moved to an archive folder shows struck through under its conversation instead of looking like a live file.

When a task calls for a visual report, the assistant writes it as an HTML page that opens right in the conversation; plain notes stay Markdown.

A reply that had already arrived could be painted over with an old error when the same message was sent again after a reconnect. The error is no longer replayed over a delivered answer.

When only the last lines of a long reply were saved to the conversation, the full text is now recovered instead of leaving just the closing line.

Files linked to a conversation could go missing or point at the wrong copy after syncing. They now keep their own change history, embedded reports are registered too, and an archived file is never silently re-pointed.

The notes list flickered while scrolling over rows with different cursors. It scrolls smoothly now.

The Slidepad's window spans the full screen height, so clicking the empty space below your notes went to Clavus instead of the app visible behind it. Those clicks now reach that app and close the panel, the same as clicking anywhere else outside it.

The new Clavus Conversation shortcut starts dictation the moment you press the Action Button, with no menu in between. When you stop, Clavus files what you said into the matching conversation, starts a new one, or shows the conversation picker. It also works with Back Tap and the Shortcuts app.

A Clavus widget puts Conversation and Transcribe one tap away and lists your most recent conversations. The small size starts a conversation dictation directly.

On iOS 18 you can add Clavus Conversation, New Chat, and Rituals as controls in Control Center, on the Lock Screen, or on the Action Button.

Press and hold the Clavus icon for New Chat, Dictate, Rituals, and a jump back into your latest conversation.

The classic grid board now scales uniformly when the keyboard area is smaller than usual, so keys keep their proportions.

The classic grid board now uses the system keyboard's key sizes, spacing, and bottom row, including ABC on the bottom left of the number board.

The keyboard's dictation card could stay hidden after a dictation ended. It now comes back when it should.

The Clavus keyboard now offers a classic grid board alongside the hex board, and AZERTY joins QWERTZ and QWERTY. Choose both in Settings.

When you ask to adjust a ritual in a chat, the assistant now knows where each change belongs: instruction and model wishes go into the ritual's instruction, schedule changes onto its schedule.

Switching the AI model a ritual runs on only lasted until the next round, then Clavus quietly changed it back. A model set directly on the ritual's schedule job now stays until the ritual's instruction explicitly asks for a different one.

A word swiped in the middle of a sentence sometimes came out with a capital letter and kept it. Autocorrect now lowercases these accidental capitals.

When the agent backend briefly stopped answering, a message sent in that moment failed outright. Clavus now checks that nothing was dispatched and retries the send once.

A ritual round that worked quietly for more than half an hour could be marked as interrupted by a restart or crash even though it was still running and finished fine. Clavus now confirms a round is really gone before settling it.

Tapping a Markdown file card in a conversation saved the file as a download instead of opening it. Cards now open the file view again; the Download button still saves.

A quick upward swipe at the start of a word could capitalize it unintentionally. The keyboard now recognizes and undoes these accidental capitals.

Summoning the assistant overlay could leave it stuck between two panels, showing one of them clipped at the edge with bare background next to it, while controls you could not see still took your clicks. It now settles and lines up on a single panel before it appears.

Clicking Download on a file in a conversation or in the file viewer did nothing in the Mac app. The file now goes to your Downloads folder and Finder shows it when it is done.

Closing the assistant overlay or Slidepad pulled the Clavus main window to the front, even when it sat on another Space. Focus now returns to the app you were in, unless you deliberately sent something over to the main window. A closed panel also stops swallowing clicks meant for the app behind it.

Tapping Download on a file, like a calendar invite or a document from a conversation, now downloads it and opens the share sheet, so you can save it to Files, AirDrop it, or open it in another app. The tap used to do nothing.

Tapping through the keyboard's language badge rewrites your text only for the language you settle on. Each intermediate step used to trigger its own rewrite.

Rounds have a stop button now, and the spinner next to a conversation comes from the server, so an active round, a running subtask or a coding session shows up on every device and survives a reload. A round used to run with no way in and nothing in the interface saying it was running at all.

Grab a card by its header to move it up or down the column, or drag a row by its checkbox or bullet to move it and everything nested under it. On touch the drag arms after a short hold, so scrolling still works.

Every card with content has a chevron next to its title. Tapping it folds the card down to its header, holding or Option clicking folds all of them at once, and a folded card shows how many open items it still has. Folded stays folded across reloads and on your other devices.

A ritual now has a language of its own, taken from your reply language unless the instruction itself is written in another one, and its learnings, its disclosure and its rounds all keep to it. Learnings used to drift into whatever language the model picked up from the topic, and because every pass reads the learnings so far, one wrong answer pulled all the later ones with it. A learning collected in the wrong language can now be deleted in the ritual history.

You no longer have to know that clicking the selected note is what deselects it.

The handoff card now carries the model and the folder the conversation is about. Writing "use Fable" into the briefing never changed anything, because a session without an explicit model ran on the machine's default, and one without a folder landed in whichever repository that engine touched last.

Cmd+X over a selected block used to do nothing at all. Cutting now removes those rows and puts them on the clipboard, and pasting them back reproduces them exactly, including prose lines, checked items and headings. A list pasted from somewhere else turns into real rows instead of a literal "- [ ]".

A save that hung used to sit in limbo, neither stored nor queued, and the next reload overwrote it with the server's older copy. Saves now give up after a few seconds and go into the queue, and an edit stranded by a reload is pushed the next time the app reaches the server.

The caret stays in the new empty row above, instead of jumping back into the text it just pushed down.

Installing from your phone used to leave Macs where Clavus starts at login with nothing running: the relaunch helper sat in the app's own process group, so macOS killed it together with the app mid-swap. The helper now survives the swap and writes every step into the updater log. The update that installs this version still runs the old helper once, so it may need one last manual start.

Long round prompts used to arrive twice, once cut off mid-sentence and once in full, and both stayed. Existing conversations are cleaned up as they load.

The note used to say pull down to refresh, but pulling down marks a conversation done and closes it, so following it archived the conversation you were trying to rescue. It now says to check your connection and resend.

A Claude session used to end the moment the model finished its turn, which killed every armed monitor and background job with it, so "I will report back when the download is done" never happened. The bubble now shows a Monitoring row while something is pending, and the follow-up lands in the same reply.

A row under iPhone settings shows what the phone is connected to and opens the setup screen, which is also where Disconnect lives.

The first message of a round used to be cut off after 2500 characters, so a follow-up often could not tell which round it was answering. Every message into a round thread now also names the round, the ritual, the run date and how the round ended.

A round whose instruction only asks for a short acknowledgement settles without arriving as a push notification. Editing a ritual's instruction outside the app no longer stacks another "You are ..." wrapper around it on every run, and an instruction that already collected several is cleaned up on its next run.

You choose Claude Code or Codex when a conversation starts, not in the middle of one. A conversation already under way hands off through a card the assistant offers, which starts the engine session with the transcript so far instead of silently switching the thread.

Coding sessions, charts, cards, copy buttons, embedded pages and the feedback loop were only ever described to it on the maintainer's own machine, so everywhere else it said it could not start a coding session even though the feature shipped. It now offers only the coding engines that are really signed in, and says so when one goes away.

A reply that arrived in full is no longer painted as failed when the transcript is compacted afterwards and that step times out.

A conversation whose messages never arrived is pushed up from the device that still has it, instead of staying empty on the server forever.

Connecting one no longer leaves the install rejecting every other model with "not allowed for agent main".

It understands the formats the assistant actually writes them in, and the intro conversation scrolls instead of cutting off at the bottom of the screen.

It opens above the keyboard on the phone instead of behind it, and units inside inline formulas read as text (15.3 °C) instead of raw LaTeX.

A new section shows whether the connected server is up to date, checks the feed on click, and installs a waiting update after a confirmation, the same way the rows for the Mac app already worked. A server this app does not manage says so instead of pretending to check.

It now opens on the screen whose hot corner you hit. With several displays it used to slide in on whichever screen had keyboard focus, with a blurred backdrop taken from the main display.

A row under iPhone settings shows what this iPhone is connected to and opens the setup screen, where Disconnect frees the app for another account or a self-hosted server. Until now a working install had no way back to that screen.

Signing in to Hosted Clavus in the browser offers to use a different account instead of silently reusing the one you signed in with last.

The setup and Server Connection windows offer a choice between Hosted Clavus and a self-hosted server. Pick Hosted, sign in with your browser, and the app finds your instance by itself and keeps the session alive, including dictation and the keyboard features. Self-hosted connections keep working exactly as before.

Pick Hosted Clavus on the setup screen and sign in with your browser; the app finds your instance by itself, and the keyboard and share sheet use the same sign-in. Self-hosted servers keep working exactly as before, now behind their own tab on the same screen.

Type a code ("pl"), an English name ("Polish") or the language's own name ("Polski"), and keep up to six. A clip spoken in one of them is written down as spoken instead of being translated into your main language.

Pasting into Mail, Notes, Notion or Word brings the picture instead of just its filename, in Safari and on the phone too, and the caret now sits where you click next to an image.

With several providers connected, your own subscription comes before the included credits, GPT 5.6 before Claude, and Opus before Fable.

Saving an edit from a panel that had been sitting in the background could delete every line written elsewhere in the meantime. The note now merges against the version your edit was actually built on.

It used to hang on "checking with the provider". The card now confirms and closes when it worked, and the login no longer narrows your install to that one model, which used to make every other model in the picker fail.

QWERTZ is only suggested when German is one of your dictation languages. The welcome message also mentions that you can write or speak in any language, and where the microphone is.

The update and connection pills no longer wash out behind the frosted strip.

A session is now compacted before a turn would grow past the model's context limit, instead of dying mid-turn. This protection is on by default for self-hosted installs.

When a run fails, the app keeps the gateway's real error message, shows it, and includes recent failures in bug reports, instead of the generic "Agent run failed".

Failed replies no longer claim you stopped them. When the gateway cancels a run on its own the error says so, and if the reply had produced nothing yet the server retries it once, so you usually just see the answer arrive.

A reply that produces nothing at all now fails after several minutes instead of spinning for hours. The check waits long enough to leave genuinely long tool calls alone.

The final paragraph was printed twice at the end of a run, which could also break a trailing question card.

It could make the app unusable until its stored data was cleared.

A screen that loads on demand retries when its first request stalls, instead of spinning forever.

It warns when the gateway is configured to reset sessions daily and recommends the setting that keeps thread history alive. The install guide carries it too.

A report from a ritual round now includes the ritual itself, its instruction, learnings and recent rounds, and every report carries your language settings, so wrong-language replies can be diagnosed from the report alone.

The card peels in the direction your fingers actually move, and the scroll-direction question is only asked where the app cannot know the answer by itself.

OpenClaw resets agent sessions daily at 04:00 by default, so on a stock install the first message each morning ran without any of the conversation's history. The assistant now detects the pending reset and restores the stored conversation into the new session.

Instructions edited outside the app, including on an unarchived ritual, are picked up reliably again.

The shell tells the web app how your trackpad scrolls, so swipe gestures match your system setting without asking you first.

It uses your assistant's actual name instead of a built-in one, and only offers coding engines that are really set up on your server.

the round's own one-line result is shown, with a clear “nothing needed from you” status — interrupted runs say so too.

they open without a connection even on a fresh install, read and write locally, and merge changes back cleanly when the connection returns.

the conversation's info sheet has a Report a bug row that takes a short description, packs the relevant logs into one package and sends it to the developer — from self-hosted installs too.

subscriptions and API keys can be connected from the web app.

warm light surfaces instead of a dark film over the glass, and a new Paper & Clay app icon throughout.

Recording the key left of the 1 — § on a Swiss keyboard — silently reset the shortcut to < instead of taking it, because that key and < share the pair of key numbers macOS uses for them. The recorder now reads the character you actually typed.

The keyboard now claims its own height at the moment the switch actually honors it.

A band of dead space had opened up below the keys — visible under the hex board while typing, and as a gap under the recording controls while dictating.

That blank state could set in when the server was unreachable at launch, or when a background refresh landed on an error page.

Each of them used to fail in its own quiet way: the dictation shortcut working only while Clavus was in front, Insert leaving text on the clipboard, a recording whose transcript came back empty.

If the app didn't return after the swap, every service stayed down until someone walked over to the Mac and started it again.

Clavus now says so instead of coming back empty-handed, and Settings → Connection test shows whether the permission is granted with a button that opens the right settings pane.

It could previously open a little smaller, with extra space around the board, depending on how it was opened.

answer each card, add free text where needed, then send the complete set together in one clear response.

Send it an event from a form, automation, or another service to start a round with that event as fresh context; Clavus queues overlapping events and lets you set a daily limit.

Useful outcomes can surface immediately, while routine upkeep stays quietly available in the Ritual history.

Parked work also resumes more reliably when results arrive later or the server has restarted.

Nothing is created until you choose Start thread, and the suggestion becomes a link back to the new conversation afterward.

Command-Z on Mac or Control-Z elsewhere now reopens the most recently archived Thread when you are not editing text.

message-like placeholders become clear topic labels, while a good existing title stays put.

Existing installations continue updating automatically.

drag for character-level control, or hold at the screen edge to remove whole words at a steady pace; moving away cleanly stops the repeat.

selected choices remain checked, attached notes appear separately, and each card presents one clear send action.

simply opening, favoriting, or renaming a Thread no longer moves it ahead of newer work.

Those lookups now run in the background and never block the interface.

cards distinguish topics and recurring rounds, results expand in place, and feedback from the parent Thread stays visible.

Markdown from any writable workspace can open directly in the built-in editor, while read-only project files remain safely viewable.

Dictation and Ask Clavus.

Every run opens as a Thread you can revisit, continue, and refine.