Changelog.

Think, at the model's own levels

• Levels, not just on and off. Think shows the levels a model lists. GLM 5.3 reads "Think · Low" and offers Low, High and Max; DeepSeek V4 Pro and Kimi K3 offer Off, Low, High and Max. Canto only ever sends levels the model lists, and remembers your level for each model.
• ⌘⇧R turns Think off and back to the level you last chose, or opens the levels for a model that can't stop thinking. Think remembers your level for each model even while it's off.
• Retry, Research this, actions, ⌘K's Ask Canto and Sparks use the Think the chip shows.

A calmer model list

• One line per model. Each model is one line: its mark, its name, and on the right only what's different: a check for the one in use, a price for OpenRouter models, or a word when something needs you (Get, Not now, Won't fit, Not running).
• Said once. Section headings say once what leaves your Mac, instead of every row.
• Details beside the list. A pane shows what the highlighted model can do: thinking, tools, images, memory, price, size and languages. In a narrow window the details fold into a line under the model.
• Recent first. Recent models head every tab, with the one you're using first.
• Retired models stay out of the list, with a note saying how many.
• One bar for how it answers. Think, Style, Reply in and the chat's memory sit in one bar under the list.

Licenses bought before 1.0

• For life. Licenses bought before 1.0 now cover 3 Macs and every new Pro feature, for life. Settings › Account says "for life" and has nothing to renew. Your Mac picks up the change the next time its license checks in.
• The license agreement changed, so Canto asks you to agree to it once more: sections 1 and 6 say the above, and section 3 lists what Ollama Cloud's model list and Contact support send.

Also in beta 4

• Help › Canto Guide opens Canto's guide on lonelyduck.io. Help › Contact Support… opens the support page with your Canto version, macOS and Mac filled in from its address; you see all of it before your message is sent, and What leaves this Mac and the license agreement list it. Contact support on a license notice or a license error opens that problem's page, and after Send Diagnostics… it brings the report's reference.
• ⇧⌘K is named for what it does: "AI commands for this note" in Settings › Keyboard.
• Ask's microphone says when dictation is off, with a button to the switch, instead of doing nothing.
• The bar that asks you to keep Canto's edit says "Review Canto's edit" and how many changes, without internal names.
• The Apple Notes import, Export workspace and archived projects use today's names and menus.
• A research note saved from an English answer full of web citations is written in English again; Canto had read each [W1] as the Polish word "w".

New figures

Answers and notes can draw ten kinds of figure from plain rows the model writes, and every table gets better on its own. Each kind is also in the "/" menu.

• Smart tables. Tables read their numbers: a bar behind each value, scaled to its column, and the best value tinted (the lowest price, cost, latency or size; the highest score, accuracy or speed). The first column names the rows and is left as written, and a column that could read either way, such as "8K Context (GB)", isn't tinted. ✓, ✗, yes, no and partial become marks. In the chat, in notes (older ones too) and on published pages. Plain table in a table's row or column menu turns it off for that table.
• Key numbers: a row of tiles, each with its source.
• Timeline: dated events on a rail, oldest first, with milestones. It reads dates however the model writes them ("2021-05", "May 2021", "Q2 2021", "the 1990s").
• Scorecard: options across, criteria down, 1–5 ratings as small meters, price levels as written, ✓ and ✗ as marks, and the recommended option tinted.
• Steps as numbered cards, each with its title and the detail under it however the model numbered or bolded the row, and a last card never left alone in its row. For and against as two balanced columns, and Layers: an architecture as layers grouped into zones. An arrow the model put between two steps or layers isn't drawn as a card of its own.
• Positioning map: options on two axes with named quadrants, every label clear of the others.
• Flow: how a total splits between stages, like requests across models or a budget across teams.
• What-if calculator: move a slider or pick an option and Canto works the result out on your Mac. Nothing in the answer runs as code, and moving a slider doesn't change the note. A published page shows the starting values.
• Explained formula: a formula with a list of what each symbol means, each symbol in its own colour in both.
• Canto checks the numbers. A figure says "Figures from the model" when its numbers aren't in the answer or its sources, and "The model's judgement" for ratings and positions. Rows Canto can't draw show as a table, never as raw text.
• One block in a note. Edit (or Return) opens a figure's rows, its kind and a live preview; Done is one undo step, and Done with no change leaves a figure as it was. Show as code turns a figure back into the code block it was, and code blocks tagged architecture, history or similar words stay code. A figure copied from a chat answer pastes into a note as a figure.
• Citations in figures are numbered with the note's References and jump to them, grouped ones such as [1, 2] at the end of a row too, and a figure's caption can cite its sources; in the chat they sit on the line instead of reaching into the one above. A list of numbers that is part of a figure, such as a formula's [0, 1] or an array in code, stays exactly as you wrote it. On a published page they follow the page's numbering, and private sources are left out.
• Everywhere a note goes. Figures export to HTML, PDF and Markdown and publish with the note; Markdown export keeps charts as ```chart blocks, and a figure inside a list or a quote no longer breaks the rest of the export. Exported HTML and PDF draw smart tables too. Search, search by meaning and the chat's reading of a note include a figure's rows, and figures in imported Markdown are stored as figures right away. In Ask's narrow panel each takes a narrow form: key numbers two across, a timeline without its descriptions, one column of steps.
• Chat answers reach for them when a question calls for one: numbers bring key numbers, dates a timeline, options a scorecard, a how-to steps.

Research notes get figures

• A figure where it helps. After a research note is written, Canto asks the model, section by section (up to five), whether a key-numbers row, a timeline, a scorecard or another figure would help, and keeps a figure only when every number in it appears in that section or its sources. The save says "Adding figures · 2 of 5" while it works.
• Your choice. Turn it off in Settings › Assistant › Research › Add figures to research notes. In Small Mac mode it stays off for models on this Mac.
• Research answers can draw them too. The research prompts now describe every figure.

After research: one card

• One card replaces four areas under a finished research answer: how many sources were read, the claims that still need a source, one Save as research note (Brief and Comparison table on its arrow) and the suggested follow-ups as chips.
• Follow-ups are suggestions you send one at a time. A chip puts its question in the message field, like any question you type. Once sent it leaves the card, and the saved note includes what it found. More… suggests new ones.
• Removed. Guided session, Compare perspectives, Expand research and running all follow-ups at once (Run follow-ups, Execute follow-ups, and the follow-up list's Act, Add and Edit) are gone. They didn't do what their names said, or ran a batch of web searches behind one button.
• Once saved, the card becomes the note: its title, how long the save took, the evidence ledger as it builds, and Open. Updating a research note and merging the update keeps the card on the note.
• Continue in a fresh chat is in the card's ⋯ menu, and shows as a key when the chat's memory is about 80% full. It keeps the research's sources, no longer saves a note first, and opens where the chat is (its page, or beside the note), as does the error card's Fresh chat.
• No second free try. Saving a research answer from its Save as note menu no longer uses a second free research try, and a research run you stopped can still be saved as a research note from that menu, free.
• Saving, sturdier. A saved research note keeps its first figure and no longer takes a figure's caption as its title. Clicking Save twice starts one save. A save in progress can be cancelled from its card, and deleting the chat stops it; either way no half-made note is left.
• Beside a chat open as a page, the inspector shows the chat's sources, every note it made (not only the last one) and what it works on, instead of an empty Ask; ⌘J and Ask go to the chat's own message field.
• Usage counts, if you share usage data. Canto counts which of the card's actions you use (save, a format, a chip, Continue in a fresh chat, Open), by name only, never what they hold. Usage data is off unless you turned it on, and nothing else new leaves this Mac.

Settings › Models

• This Mac, in one strip: your Mac, its memory in one bar, and the model limit drawn on it.
• The chat model's row is the recommendation, with four figures for this Mac: when an edit starts, a web answer, research and memory.
• Other choices that suit your Mac are listed in aligned rows, each with what it's good at and what it costs here; the one that runs Thorough research says so.
• A suggestion, once. When your chat model isn't the one recommended for this Mac, one row shows the recommended model with its numbers; Not now hides it.
• Research that takes under 5 minutes now reads "Runs well", and a Mac with little free memory shows it with one decimal instead of "0 GB free".

Fixed

• Memory suggestions stay on the chat's model. With Suggest new items on, a chat whose model isn't your default no longer unloads it after every answer to load the default one, so the next question starts at once.
• Servers get time to read. A model server on this Mac or your network gets as long as it needs to read a long prompt before its first word (at least two minutes), even one that answers at once and then reads, instead of being stopped after 45 seconds.
• Quicker research steps. Local research with many tool steps starts each next step sooner: the shared part of the prompt is no longer read again.
• OpenRouter fetches its model list once when you connect instead of many times at the same moment.
• Search by meaning. Quitting or updating Canto while it indexes a note no longer counts against that note. Notes it skipped after repeated failures get another try with each new version, and Settings › Advanced › Diagnostics lists them with Try again.
• "Finishing the upgrade" no longer stays on the search index when it can't finish, and a catch-up after an update says "Updating the search index".
• Citation markers stay. A new research note keeps citation markers that match no source, as written, and Note health lists them; nothing is deleted from your text.
• Palettes. Canto itself checks a palette other than Canto against your plan, not only the Appearance page; the palette follows your vault to your other Macs with iCloud sync; and the palette your 0.x theme had is kept when you first open 1.0.
• The Apple services helper. After it stops on its own, memory figures no longer start system tools every five seconds, the Apple Intelligence mark comes back, and the Helpers page shows the restart it just did. Helpers now says "If it stops more than 5 times in 5 minutes", which is when Canto waits for Restart.
• Chart data from a table. Binding a chart to a table no longer reads ranges such as "5–8" as a number.
• "Created note" keeps a long note title on one line.
• Stopping a model on this Mac no longer leaves Canto waiting for a word that never comes, and cancelling a research note while a server, cloud or OpenRouter model writes it says it stopped, not that it failed.
• Tab with a citation, inline math or an image in a list selected indents the item again.
• Published pages in dark mode draw chart colours, formula terms and layer names in the lighter dark-mode colours.
• Upgrading from Canto 0.8 or earlier keeps your theme, text font and sidebar rows.
• Long-form drafts written by an earlier beta upgrade their formatting again when opened, and resuming one no longer pauses as if you had edited it.

Lighter while it waits

• Memory is read without starting a process. Canto used to start three small programs every second, for as long as it ran, to read the Mac's memory (one of them made Canto's main process wait each time). Canto's own helper reads it now, every 5 seconds, and not at all while every Canto window is hidden. The Memory window and the parallel-answers setting still update every second while they're open. A change of model still shows at once.
• Linked folders stay quiet. With a folder linked in Folders on this Mac, Canto's search re-indexed every file in it every 3 minutes, even when nothing had changed. It now indexes only what changed.
• Calmer figures in Settings › Models. "Free now" and the memory bar move in whole gigabytes, only when memory really changes.

Updates say what they do

• Finishing the upgrade. The first launch after an update can have one-time work to do: backing up and updating your notes database, turning notebooks into notes, clearing data the old sync left behind, building the search index. While it runs, the status dot says Finishing the upgrade, and its panel lists each step with its figures ("41 of 72 sources") and when it's done. All of it happens on this Mac.
• A quicker first launch after an update. The backup Canto makes before updating your notes database, and before turning on iCloud sync, is now an instant copy on your Mac's disk (an APFS clone). Before, Canto wrote the whole database again, which on a large vault was hundreds of megabytes.

Settings

• Advanced › Helpers. Shows whether the Apple services helper is running, how much memory it uses, and whether it stopped on its own this session. Restart starts a fresh one without quitting Canto. If a meeting recording or a transcription is running, Canto asks first.

A new look

• One toolbar with the ⌘K field, a status dot (green when all is well; click it to see what Canto is doing) and Ask. It fits any window width, and says when the chat you're looking at runs on a cloud model.
• A grouped sidebar: Home, Search, Inbox and Today; your pinned notes; the notes tree; recent chats; projects; the library. It never swaps itself out: All chats, Attachments and linked folders open as pages, a tag opens a page of its notes, projects open to their chats, and Meetings, Remembered (0.10's Continuity, now a calmer list) and the vault graph live in the Library. Every group folds and stays as you left it.
• No tabs. Each pane shows one document, with back and forward (⌘[ and ⌘]). Split up to three panes side by side; they share the window evenly. Word counts and the like moved to the inspector's Info.
• Home (⇧⌘H) is one calm column: today's note, what you opened last, what's waiting in the Inbox, and the Spark when there is one.
• Glass and density. Settings › Appearance has Glass (Full, Subtle, Off) and Compact or Comfortable spacing. Canto turns the glass off when macOS's Reduce transparency is on.
• Four palettes. Canto's own is free; Mist, Sepia and Graphite are part of Pro.
• The menu bar has a Go menu, and items for Ask, Wide Chat, the Inspector and Note History. Open Data Folder opens the folder Canto really uses.
• Canto remembers where you left its window, and brings it back onto your screen if that monitor is gone.
• The status dot's panel speaks plainly: one memory figure, the model's state in words, search by meaning with Turn on or Try again, and Details in Settings.
• Tags can be named in Korean, Japanese or any other language (a name made only of emoji is refused, and Canto says why). A tag's page can rename the tag, change its colour or delete it; deleting a tag never deletes its notes.

⌘K for everything

⌘K opens one bar. Type to go to a note, chat, project, folder or Settings page, to run any command, or press ⌘↵ to ask Canto. Search looks at words and meaning together, across notes, chats, canvases, projects, folders and the text inside attachments, and finds Korean words inside longer ones. It also finds files in Folders on this Mac by name and by what they say, and opens them. Note links in results show as their titles, and an action you name ("switch to dark") comes first. It replaces ⌘P; ⇧⌘F opens it ready to search.

Ask, beside your note, with sources

The inspector on the right speaks for the document in front of you. ⌘J opens Ask.

• One composer everywhere. Ask beside a note, a chat page and Home share one glass composer in one calm row: + and Sources (what Canto reads, the web, past chats, Research instead) on the left; the model, with Think beside it, and one button on the right that is the mic when the field is empty, Send, Research when research is armed, and Stop while Canto works. It stays one row at any width; words give way to icons as it narrows. Beside a note it rests as a slim bar until you click it or press ⌘J.
• Choose what it reads: this note, your selection, its folder or project, or everything, and the web only if you turn it on. With Web on, Ask may search on any message and decides when to.
• Ask edits the note beside it when you ask for a change ("check off the first task", "add a summary at the top", "clean up the formatting"); a question about the note is answered in the chat. The line under the composer says so: "Can edit “Launch plan” with your review".
• Edits cite their sources. When Ask adds to a note, its citations become numbered chips as the edit lands, from every source the chat has found, and a number like [11] cites that reference of the note again. References updates itself, all in the same review and one Undo. Under the answer, "Edited this note" says what changed and what linked, with Show in note and Undo. A citation that matches no source stays in the text and shows in Note health.
• Turn a chat into a note however you say it ("make this a note", "save this answer", "이 내용을 노트로 만들어줘"); once a chat has an answer, Ask can always save it. Ask for a research note and it gets a cover, citations, References and an evidence ledger.
• Sources you can check. An answer about your notes cites the passage it used, and a click opens it. Answers about your vault come in one step, much faster than in 0.10. Sources under an answer are one row that opens to the full list, cited sources first; point at a citation like [W3] to see its source.
• The model list floats over Ask, with tabs for this Mac, your servers and the cloud, and shows each model's maker, build and what it can do. Choosing a model never loads it or starts a download: loading happens when you send, and the chat says so. Models you don't have yet open Settings › Models.
• Think can be turned on before the first message, and Canto remembers it for each model. Research instead (⌘↩, or in Sources) researches one message without changing the chat or turning on the web; it shows as a flag in the composer with its options. Under a long web answer, Make a research report and Suggest follow-ups are one click away.
• Memory. A ring around the model's mark shows how full the chat's memory is. Its menu shows the numbers, what's loaded now, what to do when it fills, and how the model answers: Think, Style and Reply in.
• While Canto works, one line says what it is doing, with a live count of tokens and the time so far: "Reading the results · 54% · 31,600 of 58,400 tokens". When a model reads slowly, Canto offers a faster one for the chat; after a minute of thinking on this Mac, Write now ends the reasoning and starts the answer.
• A stopped answer says so: how long it ran, whether the note changed, and Try again (or Try again without Think).
• A failed answer is one card with buttons that fix it ("Write from the top 8 sources", "Save sources as a note"). An empty answer names the model and says only what Canto knows; a chat that no longer fits says so and offers to compact older messages or start a new chat; a cut-off answer says whether the chat filled the window or your Longest answer setting stopped it.
• Steps read clearly: searches show their query, pages their sites, edits link the note, and back-to-back thoughts are one row. How I got this says what was searched and read in plain words. Every answer names the model, its build and the style it used.
• Under each answer, Copy and Save as note are labelled (Save as note › As a research note turns a well-sourced answer into a research note: Pro, or one of the free research tries); Export, Remember and Regenerate are under ⋯.
• Before the first message to a cloud model, the chat shows what will be sent, to whom and about what it costs. While a chat's model leaves this Mac, one line inside the composer says so; ✕ hides it for that chat and model, and "Don't ask again for this chat" is remembered too. Each cloud answer says where it ran beside its time; what it cost and this month's total are one click away.
• Context shows what links to the note (with the sentence), notes related by meaning, its sources and its outline. Info shows its properties, counts, versions and whether it's published.
• Size. Ask opens at about a third of the window (wider on big screens, never squeezing your note) and re-fits when you resize. The edge beside it is easy to grab, shows where it snaps and works with the arrow keys; double-click or Return goes back to the default. On small windows the sidebar steps aside before Ask gets cramped. Wide chat (⇧⌘J) gives Ask the window while your note stays beside it in a reading column; Esc goes back.
• Keys in the composer: ⌘⇧P model, ⌘⇧R Think, ⌘U attach, ↑ edit your last message, Esc stop.

Research notes

• Saved at once. A research note opens as soon as it's saved. Its evidence ledger is built in the background, with progress under the note's title and under "Created note" in the chat.
• The note keeps up. After Ask edits a research note, Canto checks the new sources and updates the ledger for only the sections that changed, each step shown under the title. It runs on the chat's model; if that model isn't available, the header says the ledger is behind and offers Update.
• Refresh references (in ⋯, Note health and the References panel) turns citation markers already in a note into linked chips.
• In your language. "Write the report in Korean" applies to the research and the note, its headings, References and Evidence included. Korean words stay whole at line ends in notes and chat answers, and spell checking is off in research notes and no longer underlines Korean.
• Note health says what happened when a ledger can't be built: the model kept thinking, was cut off, answered in the wrong format, found no claims, or couldn't be reached. "Build again" runs on the research chat's model and says whether it leaves this Mac. Without a ledger, the header counts the sources cited instead of claims.
• Only the answer is read. Plans, citations and the ledger never read a model's reasoning. A model that only thinks is stopped and Canto says why, instead of asking it the same thing again; long notes are read in parts.
• No folded copy of the chat's answer. Research notes no longer carry 0.10's "Original interpretation (verbatim answer)" block, and the setting that kept it is gone. Notes that have one keep it, hidden, and leave it out of exports and published pages.
• "Follow-up findings" appears only when a follow-up ran, and suggested follow-up tasks written in Korean are kept.

Writing

• Opening, reading or pinning a note never marks it edited, moves it to the top of your lists, or adds a version.
• If Canto can't save a note, it says so on the note ("Not saved · Retry") and keeps your edits until it can.
• Typing shows on the next frame, even in long notes.
• The "/" and [[ menus are glass menus in your palette, grouped in plain words, with the Markdown shortcut beside each block type. Table inserts a 3 × 3 table at once. [[ shows each note's folder and offers "New note “…”" when nothing matches.
• Find in a note is a small floating field that doesn't push the note down.
• Smart quotes and dashes leave LaTeX alone while you type it, and in notes Ask writes.
• Math in table cells renders ($\alpha$, $x^2$), while pricing tables keep $, $$ and $5 as text. Notes from 0.10 are fixed when you open them, and saved with your next edit.
• Editing a note with Ask no longer rewrites its References, and References damaged by 0.10 are repaired when you open the note. Ask can't drop a paragraph's citations, links or note links when it rewrites it.
• Section roles show only in research notes, once per section.
• Notes created by other apps or windows appear in the sidebar right away.
• Links, note links and link chips use your palette's link colour, in light and dark. Callouts, inline code and Mermaid diagrams follow your palette; diagrams open at their natural size and show their source only while you edit them.

Inbox and quick capture

Press ⌥Space in any app: a small window takes a thought and saves it to the Inbox, without switching to Canto. The Inbox also collects suggestions Canto makes on this Mac (tags, links, notes to tidy), which you accept, edit or dismiss. "Suggest tags" stays until the note has tags, and says so when no model is loaded.

Settings, in its own window

⌘, opens Settings as a window, one size (1000 × 740) like a Mac preferences window, with fourteen pages and a search field. It opens where you left it, on the page you were on.

• Every page uses the same rows, switches, pop-ups and keys, in sentence case. A row that depends on another switch is greyed out and says which one.
• Search finds the lines inside panels (Usage data, Back up now, a shortcut), forgives a typo, and lands on the line with keyboard focus. ↑ and ↓ move through the sidebar.
• Keyboard is its own page: change a shortcut by typing it, in Mac symbols (⇧⌘K); a shortcut that is taken is offered as a swap. Every other shortcut is listed by group.
• About and Updates have their own pages under Canto, at the end of the sidebar; Account holds only your license. Canto › Check for Updates… opens Updates and checks. Download automatically waits for Check automatically; What's new is on About. An update says so in a toast at the bottom of the window, with Restart, and the sidebar marks Updates. A check that fails says so in plain words.
• Version history can be turned off; Show in Finder works wherever your vault is; Save edits back to linked files can be turned off.
• Your chosen model is kept in your vault, so it survives a reset of the app's browser storage.

Privacy and search

• Memory & privacy › What leaves this Mac lists everything that can send data (web search, cloud models, iCloud sync, publishing, license checks, update checks, usage data, site icons, pictures from the web, other apps connected over MCP, pages you link to), what each one sends, where and when, and a switch for each. When a switch is off, Canto doesn't send it.
• Search by meaning runs entirely inside Canto: nothing is fetched from the web to start it, and it works offline. It is on by default; if it was off in your vault, it is turned on once and Canto tells you. Its status is the same everywhere: off, starting, indexing, up to date or needs attention, with Turn on or Try again.
• Cloud models is one switch, with OpenRouter, Ollama Cloud and each server elsewhere under it. If you already used Ollama Cloud models, they stay on; turning Ollama Cloud off keeps your Ollama models on this Mac working.
• Pictures from the web. Pictures in notes that point at a website load from that site; turn it off and they show a placeholder with a Load button instead, and nothing is fetched.
• Web source icons no longer go through Google. With Site icons on (off by default) they come from each site; off, sources show a plain globe.
• Turning Web search off also stops links you paste into a chat from being opened. A link whose site doesn't answer gives up after 20 seconds.
• Settings asks Tavily for your usage only when you press Check usage.

Models for your Mac

• A new lineup: Qwen 3.5 4B and 9B, Granite 4.2 3B, Gemma 4 12B and 26B A4B, Qwen 3.8 27B and Qwen 3.6 35B A3B, with larger ones under Advanced.
• Recommended for this Mac. Canto picks a model by your Mac's chip as well as its memory, and says how long to expect: "about 35 seconds before an edit starts, about 3 minutes for 10 sources". 8 GB Macs, the MacBook Neo included, get a Small Mac setup with Qwen 3.5 4B; 16 GB Macs get Qwen 3.5 9B (or the 4B where it starts edits faster), with Gemma 4 12B for better writing; 32 GB and up get the fast Qwen 3.6 35B A3B, with Qwen 3.8 27B one click away.
• Builds. Every model shows its build (Compact, Balanced or High precision), and the builds of one model are one row; a popover shows which you have and which suits your Mac. A High precision build of Qwen 3.6 35B A3B is there for Macs with 64 GB or more.
• Chats size their memory to the question. 0.10 reserved room for the longest possible chat every time; now a chat starts small and grows, and the model's memory is stored at 8 bits. With the 9B model, a chat used about 15 GB of graphics memory in 0.10 and about 6 GB now.
• Only the tools a question needs are offered to the model, so simple questions start answering in under a second instead of several.
• Quick helper. Where Apple Intelligence is available, it writes titles, tags and short summaries with no download.
• Settings › Models is one page: what your Mac is and what Canto recommends for it, what each job uses, every built-in model in one list, your servers and cloud models, Apple's services and voice. Each model is drawn the same way, with its maker's mark (Settings › About says the marks belong to their owners).
• Memory is a number with a coloured dot (fits, not right now, won't fit), and Settings and the model list agree on it: a model that only needs more free memory stays available with a note; one that can't run on your Mac says so. A bar only means a download, which shows how much, how fast and how long, and can be cancelled. Downloads check free disk space first.
• This Mac shows your chip, memory, how much macOS lets models use, and what each job takes here; on 8 GB Macs it explains Small Mac mode.
• Apple on this Mac shows Apple Intelligence, Apple Vision and Apple speech with their state, languages and help, and two switches: use Apple Intelligence for titles, tags and summaries, and describe images.
• Voice is rebuilt: the Korean-tuned dictation model comes first when you dictate in Korean, and the microphone row asks only when macOS can still ask.
• Canto tells you when a newer upload of a model you downloaded is available.
• Usage data, if you turn it on, includes your chip tier (such as M4 Pro) and model load failures by memory range.

The AI engine

• Ten or more web searches fit on every Mac. For research and web chats on a Mac's own model, Canto plans the model's memory before it starts, reads each web result as a short digest, and condenses older results instead of overflowing.
• On Qwen 3.5, 3.6 and 3.8, each tool step reads only what is new instead of the whole chat again.
• A follow-up on this Mac starts sooner: Canto keeps what the chat already read and reads only what changed (about 4 times faster to the first word on Qwen 3.5 9B).
• An edit on this Mac thinks briefly, then writes the change straight into the note.
• Stop works at once, even while the model is reading a long prompt.
• The chat says what the model is doing: "Reading the search results · 40% · about 2 min left", "Loading Qwen 3.5 9B · 40%".
• Long notes are read section by section, so "find all errors in this note" sees the whole note.
• A failed answer shows one plain message once, and when sources were found Canto writes from the strongest of them instead of stopping.
• Macs with 8 GB start with AI on, in Small Mac mode. Safe Mode is only for recovering from a launch that didn't finish, and says why it started.
• Gemma 4 models get the context length they can actually hold on 16 GB Macs.
• On battery, in Low Power Mode or when the Mac is hot, long answers use a faster model and pause between steps, and say so.
• AI edits to a note apply completely or not at all.
• Background jobs (classification, compaction, slash commands, tag suggestions) run only on models built into Canto, never on a server or cloud model.

Servers and cloud models

• Servers are read, never probed. Adding a model server never loads its models or sends them test prompts: Canto reads what the server reports. Ollama, LM Studio, oMLX and llama.cpp's server each get their own connection, with their state, version, what is loaded, and plain help when one isn't running or wants a key. Settings › Models finds them on this Mac when it opens (it only asks, and nothing leaves the Mac); you can turn that off in Advanced.
• Context that fits. Cloud models (Ollama Cloud, OpenRouter) get their full context, up to 262,144 tokens, and use no memory on this Mac. Models served from this Mac are sized to each chat and the memory free, the way Canto sizes its own, growing at most twice per chat. Answers always get room to finish (at least 2,048 tokens, 4,096 with Think). Canto unloads only the models it loaded, when you quit or after a few idle minutes, and never a cloud model. ⋯ › Keep the server's context setting turns sizing off per server.
• Think off reaches LM Studio, oMLX and llama.cpp's server, and thinking levels follow what each model offers. Ollama models you added in Canto 0.10 (Kimi K3 Cloud among them) are listed again once, so Think off works for them too.
• Servers show their models grouped (on this Mac, Ollama Cloud, not for chat), five at a time. Models that can't chat (embedding and classifier models) aren't offered for chat. Ollama Cloud models are recognised by what Ollama says about them, so a renamed copy is still labelled as cloud.
• When the server your chat model runs on isn't running at launch, Canto keeps your choice and uses a built-in model until it's back, and says so.
• No part of Canto's window can read a server's saved API key.
• New: OpenRouter, for bigger models you rent by the answer on your own account. Off until you connect it, in one sheet that says what leaves your Mac, who sees it and what it costs. Only hosts that keep nothing by default; prices before you send; a monthly limit ($10 unless you change it) kept on this Mac; never used for background work. Free on every plan.

Images and scanned PDFs, read on this Mac

macOS's own Vision framework now reads the text in images and scanned PDFs and recognises what a picture shows, so search and Ask can use them. On macOS 27 with Apple Intelligence on, each image also gets a short description. Nothing is uploaded, and a chip under each image says how it was read. Scanned PDFs are read up to 500 pages (0.10 stopped at 15).

Before macOS 27, Canto reads the text and labels in images but not what a chart or photo shows, so questions about charts and photos are less precise until you update. The separate vision model is gone; if you downloaded it, Settings offers to reclaim the space.

Meetings (Pro)

Record a meeting on this Mac: your microphone is "You", and your Mac's own sound (the other people on a call) is "Others". Canto transcribes as it goes, on this Mac, then writes a note with a summary, decisions and action items using your local model. If Canto quits during a meeting, it offers to finish it at the next launch. Nothing leaves your Mac. Meetings live in the sidebar's Library, and each recording starts from a sheet.

Transcripts need macOS 26 or later; on macOS 14 and 15 the recording is attached without a transcript. Recording your Mac's sound needs macOS 14.4 or later.

Sync through your own iCloud (Pro · 3 Macs)

Settings › Sync & storage turns on sync through your own iCloud account. Every note, folder, tag, chat, attachment and synced setting is encrypted on this Mac, with a key only your Macs have, before it leaves; Apple can't read it, and LonelyDuck never sees it. Your first Mac asks you to choose a recovery passphrase; your other Macs get the key from iCloud Keychain or ask for that passphrase. Canto backs up your notes before the first sync. When two Macs change the same note at once, both versions are kept. Diagnostics reports say how sync is doing.

The old iCloud Drive sync is gone (see below). In this beta, Canto checks for changes when you switch to it, when your Mac wakes and every 5 minutes, so a change can take a few minutes to arrive.

Vault graph

• A living map of your notes that settles, drifts gently for half a minute after you touch it, then rests. Drag a note and its links follow; hover to see its neighbourhood; folders are soft groups; notes that mean the same thing are joined by a dashed violet line.
• Click a note to see what it links to and what it resembles; Enter or Open note opens it in the same pane, and ⌘[ brings the graph back. ⌥-click opens it beside.
• Find a note lights up the matches and flies to the one you choose. Show Links, Similar, Tags and Folders fade in and out without rearranging the map.
• Large vaults stay smooth: 5,000 notes open in a moment and pan and zoom at the display's full rate.
• The graph opens where you left it, and works from the keyboard (arrow keys walk the links) or as a list. With Reduce Motion it lays itself out before it appears and stays still.
• Similar notes are found on this Mac by search by meaning; nothing leaves the Mac. Notes changed by other apps, the Inbox, meetings, imports or your other Macs show up in an open graph.

Canvases

• Ask works on canvases. Beside a canvas, Ask reads its cards and, once you say yes, adds cards, puts your notes on it, links the related ones, tidies it or draws a mind map. Everything one answer changed is one step: Undo under the answer (or ⌘Z on the canvas) takes it back.
• "Link related notes" on a canvas's ✨ and right-click menus joins note cards whose notes are alike, without a chat.
• Canvases look like the rest of Canto: a paper board, glass tools, sentence-case titles ("Untitled canvas"), and sticky notes that stay readable in dark mode.

Actions & styles

• Commands, skills and personas are now Actions & styles, with their own page in Settings. Actions do one job on what you point at (selected text, the note or a chat); styles change how Canto answers.
• Saved and imported actions are never deleted on launch. If earlier versions left copies of built-in actions, Canto asks once what to keep.
• "/" in Ask offers the note's actions as well as the chat's, and Return runs the highlighted one. Tab puts an action in the message so you can add words; the chat shows the action, not its instructions.
• The editor's "/" starts with "AI · this note"; the selection bar shows your pinned actions and "⋯" for the rest; ⇧⌘K lists AI actions for the note, the selection and the chat.
• Built-in actions can be switched off or customized, and reset to the original. Shorten, Summary at the top and Key points and to-dos replace the overlapping ones; Find a source and Translate join the selection actions; "Open a random note", "Surprise me" and "Send to future me" moved to ⌘K.
• Making an action is one sheet: what it works on, instructions in plain words with values like "Selected text", where the result goes, and Try it on one of your notes. Actions can run in steps, and ask before a step.
• Styles: a default for new chats, Creativity instead of temperature, and a style made from a sample of your writing.
• Import shows what each item will do before anything is saved, and never doubles what you have.

Trash and Library

• The Trash is a place beside your notes, not a dialog: search it, filter by kind, select several, and see each item in the inspector before you restore it or delete it for good. Every row says where it came from, how big it is and how many days it has left; if its folder is gone, it says where a restore will put it.
• Deleted notes, canvases and chats are deleted for good after 30 days. Choose 7, 30 or 90 days, or never, in Settings › Sync & storage › Trash.
• Deleting for good asks first, naming what goes and what else changes (links in other notes become plain text), and says that backups made before keep a copy until they age out. Deleting a note for good also removes its own chats.
• A note, canvas or chat in the Trash shows one quiet bar with Restore and Delete forever…, and a chat in the Trash has no composer.
• Right-click menus work from the keyboard, open where you click and speak in sentence case.
• Move to… is a quick sheet: type to filter folders, arrow keys and Return to move, or make a new folder there.
• Show info opens the inspector's Info beside the note. Linked folders can be renamed right in the sidebar.
• Exports start with the note's title as a heading; Markdown import keeps checklists, tags and note links; Version history shows changes paragraph by paragraph.
• A file in Folders on this Mac that can't be opened (a damaged or password-protected PDF) says so, and why.

Throughout

• Dialogs are Canto sheets: a light frosted backdrop, a soft spring in and out, and keys that say what Return and Esc do. Return presses the highlighted key; Esc or a click outside closes; focus goes back where it was. There are no system alert boxes: every question and error is a Canto sheet or a short message.
• Before something is deleted for good, Canto names what goes and starts on Cancel; ⌘⌫ confirms.
• Buttons are one family of Mac keys throughout, with the arrow cursor, the same focus ring, and key hints once you use the keyboard.
• Toasts are capsules at the bottom of the note, never over a composer; "Undo ⌘Z" works with the keyboard.
• When something you did doesn't work (a rename, a pin, a move, blocking a site), Canto says so and puts the old state back.
• One set of icons, drawn in the same weight. Text size and density reach every panel; nothing in the app is smaller than 10.5 pt. Counts and times use even-width figures.
• Red, amber and green only mean something is wrong, waiting for you or done, in every palette and in Increase Contrast. Highlighted text is readable in dark mode.
• Every button can be named by VoiceOver; collapsible sections, diagrams, tags and the sidebar's selection boxes work from the keyboard; Reduce Motion stops decorative animation everywhere.
• Menus and popovers stay inside the window and open upward when there's no room; floating bars, menus and sheets always stack the same way.

Free and Pro

Canto is free on one Mac with nothing to unlock: the whole app offline and encrypted, unlimited AI on this Mac (Ask with citations, quick research, rewrite, summarise), any local model, Ollama, LM Studio and other model servers, OpenRouter on your own account, dictation, image descriptions, agent edits to one note at a time, canvases, graphs, version history, backups, export, and 3 published pages. A free copy sends nothing about a license.

Canto Pro adds deep research reports with an evidence ledger, meeting capture, agent edits across many notes, write access for apps connected over MCP, all palettes, and 50 published pages with your own domain. Pro · 3 Macs adds sync between your Macs. Pro · 1 Mac is $29 (₩39,000) and Pro · 3 Macs is $39 (₩49,000), once, with 12 months of new features that you keep forever; another 12 months is $19 ($15 before the first ones end). Each Pro feature on this Mac can be tried three times first (except sync and publishing).

• Pro sheets show the feature working: a short preview, three plain lines, and both prices side by side. Return tries it while you have free tries, and buys once they're used.
• One license form. "I have a license" turns the Pro sheet into the license form right there: paste the license or the whole email, see what it is (checked on this Mac, nothing sent), then activate. The same form is in Settings › Account, on the link from your email, and in ⌘K › Activate Canto Pro.

If you bought Canto before 1.0, your license moves over by itself the first time the beta reaches LonelyDuck, now covers 3 Macs, and includes 12 months of new features from this release. After that, a Mac with Pro checks in with LonelyDuck about twice a week (your license and an id for this Mac, never notes), instead of sending your license key every few minutes.

Publishing

Free accounts can publish 3 pages and Pro accounts 50, with a custom domain for Pro (Settings › Connections). Pages you published before 1.0 stay up and stay editable even if you have more than 3; only new pages count against the limit.

• A new page, made for the web. An outline that follows where you are, the source behind each citation in the margin (a sheet on a phone), the evidence ledger with filters, Copy link, a Markdown copy and Copy for AI, and a link preview drawn from the note. Pages follow the reader's light or dark appearance in your palette, with a switch, and Korean text breaks properly.
• Pages you published before keep their look until you update them; they no longer show the folded copy of the answer, and their headings are readable in dark mode.
• The publish sheet counts sources, not links, and can put your name on the page (optional, off by default). Math in tables and diagrams appear on the page as they do in the note.

What was removed, and why

• Notebooks are notes now. Markdown cells became text, code cells became code blocks, each cell's last output became an "Output" block, and images a cell produced became attachments. Folders, pins, tags, dates, projects and chat references came along. Before converting, Canto saved every notebook as a .cnb file and as Markdown in your Backups folder ("Notebooks (before 1.0)"). Running code inside Canto is gone with them, and so is the Python runtime; if you used it, Canto offers once to delete the packages it had installed. Running arbitrary code inside your notes app was a security risk we no longer want to carry.
• Cloud speech-to-text (Deepgram, AssemblyAI) and a hidden browser speech engine that sent audio to Google. Dictation runs on this Mac only.
• MCP automation (screenshots, synthetic clicks and window control through canto_invoke). It let any connected app drive Canto's window. The vault tools stay. If you had automation on, MCP settings says so once.
• Turbo mode and the bundled model server. One engine runs local models again, with less memory than Turbo needed. If you downloaded Turbo files, Settings shows how much space you can reclaim.
• The old iCloud Drive sync. It kept every edit forever in your database (a 123-note vault weighed up to 850 MB) and could lose changes. Sync through iCloud replaces it. Attachments that existed only in iCloud Drive were copied back to this Mac first, and the old history was cleared, which makes the database much smaller. Once sync through iCloud has finished, Sync & storage offers to move the old iCloud Drive/Canto/Sync folder to the Trash.
• Vault Manager and the document agent panel. Ask beside a note or a canvas does their work.
• WebM, MKV and AVI files. Audio and video are now read by macOS itself instead of a bundled FFmpeg, an Intel-only build that ran through Rosetta. Canto reads MP4, MOV, M4V, M4A, MP3, WAV, AAC, CAF, AIFF and FLAC, and says plainly when it can't open a file.
• The free AI trial of cloud features, replaced by Free and Pro.

The app is also much smaller, because the model server, the Python runtime, the code editor and FFmpeg no longer ship with it.

What happens once, after you install

• macOS treats Canto as a new app. Canto 1.0 is signed as io.lonelyduck.canto (0.10 was com.canto.app). The first time you dictate or record a meeting, macOS asks again for the microphone; the first Apple Notes refresh asks again for Automation; if your Canto data folder is in Documents or Desktop, macOS may ask once for that folder; and macOS may ask once about an item Canto saved in your Keychain (choose Always Allow). Your notes and settings carry over unchanged.
• The license agreement, once. Canto 1.0's agreement is rewritten (what can leave your Mac, AI answers, Free and Pro), so the first launch asks you to read and agree to it before anything else.
• Apps connected over MCP need their setup again. The MCP server now always asks for its access key, also from apps on this Mac. Copy the setup for each app again from Settings › Connections › Other apps.
• The 1.0 sheet shows once, with a link back to this page.

Known issues in this beta

• Sync checks for changes every few minutes rather than instantly.
• Image descriptions need macOS 27 with Apple Intelligence (see above); videos you added before updating to 27 are described after you index them again. Meeting transcripts need macOS 26, show no live word-by-word text while you record, and don't tell the other people on a call apart.
• Search by meaning understands English only; Korean is found by words. Attachments, canvases and projects are found by words.
• With a model on this Mac, each follow-up message reads the whole chat again, so long chats take longer to start answering.
• A published page's link preview image leaves out Japanese and Chinese titles.
• OpenRouter's sign-in page names localhost instead of Canto. Direct Anthropic, OpenAI and Google keys aren't in this beta.
• A deleted folder isn't listed in the Trash; its notes are, and restore to the nearest folder that still exists (each row says where).
• Chats from 0.10's canvas agent aren't listed in Ask's history; they stay in your vault, and chat recall still finds them.
• If your only Mac on a Pro · 1 Mac license is gone, write to us to release it.
• This beta was tested on macOS 27. If you use macOS 14, 15 or 26, please tell us what you find.

Tell us what you find: About › Send diagnostics shows you everything it would send before it sends anything.

Going back to 0.10.3

The beta changes your notes database the first time it opens it (to version 62), and saves a byte-for-byte copy from just before, in the Backups folder inside your Canto data folder, named pre-migration-v58-<date>.db (the number is your previous version's). Canto 0.10.3 is on the Canto releases page.

To go back to exactly where you were (anything you wrote in the beta is not in that copy):

1. In the beta, export what you wrote or changed since installing it (File › Export To… › Export as Markdown).
2. Quit Canto.
3. Open your data folder: in Finder, Go › Go to Folder… and type ~/Library/Application Support/Canto (or the folder in Settings › Sync & storage › Data location if you moved it).
4. Move canto.db, and canto.db-wal and canto.db-shm if they are there, into another folder. Don't delete them yet.
5. Copy Backups/pre-migration-v58-<date>.db into the data folder and rename the copy to canto.db.
6. Install Canto 0.10.3 in place of the beta and open it. It may ask for your license key again; it's in your purchase email.

If you open 0.10.3 without restoring the copy, it opens the beta's database and everything you wrote in the beta is there, but each former notebook appears twice (as the notebook, and as the note the beta made of it), the old iCloud Drive sync is off, and it may ask for your license key again. Turn the old sync back on only if you didn't move the iCloud Drive/Canto/Sync folder to the Trash in the beta. If you later install the beta again, notes you wrote in 0.10.3 come along, but changes to notebooks made in 0.10.3 don't.

If you turned on sync through iCloud in the beta, your notes stay in your iCloud; 0.10.3 doesn't read them.

Close this tab whenever you're ready — it won't appear again.

What Canto Sends, and What It No Longer Does

• The installation check-in is gone. Until now, Canto sent a small message to our server at launch and every 5 minutes: an installation ID, the app and macOS versions, and your license key if you had one. It ran even if you had turned usage statistics off, and nothing in Canto told you about it. That was wrong, and it's removed.
• License checks every 6 hours, not every 10 minutes. Canto still checks, so a refunded or revoked license stops working, but it asks far less often.
• Activation asks less. Activating a license no longer looks up your public IP address or your city through third-party services, and license requests no longer include your Mac's name or macOS build. On its next license check, Canto replaces the Mac name we had stored with "Mac".
• Update checks every 6 hours, plus once at launch, instead of every 5 minutes.
• Usage statistics can't be tied to your license. If you opted in, they no longer carry a value derived from your license key.

What still leaves your Mac: license checks (only if you have a license); without a license, a check of how many free AI tries are left (an id for this Mac, the app version and the platform, when Canto opens and when your Mac comes back online); update checks; and the things you turn on or send yourself — Sync, Publishing, web search, cloud models, diagnostics reports, and usage statistics if you opted in.

Chat Receipts Tell the Truth About Cloud Models

The context receipt under a Canto Chat answer said Privacy: Local only whenever no web search ran — even when the answer came from a cloud model, such as an Ollama cloud model. In that case your question, and the notes Canto read for it, went to that service.

The receipt now says where the answer was made: Local only, Cloud model, Model on your network, or Model on a remote server, and whether the web was used. Answers from before this update are relabeled too.

Undo for Delete

The trash icon on a sidebar item moved it to the Trash in one click, and for notes and notebooks said nothing at all. Now Canto says Note moved to Trash with the note's title and an Undo button, and the same for notebooks, canvases and chats. Deleted items still go to the Trash, where they stay until you empty it.

The Trash window now closes with Esc.

Security Fixes

• Keychain. Saving a key or password in the Keychain no longer passes it through a command shell, so a value with special characters can't run a command, and secrets no longer show up in the list of running processes. Your existing Keychain items stay where they are, and you won't see new Keychain prompts.
• Development settings are ignored in the installed app. Settings meant for development, such as loading the interface from a local test server, could change how the installed app behaved if another program launched it that way. The installed app now ignores them.
• MCP automation needs your token. If you turned on the MCP automation tools, the installed app ran them without checking the access token. It now requires the token set in Settings.
• Your database key is kept. If Canto couldn't find your database where it expected — for example, a data folder on a disk that wasn't connected — it replaced the database's encryption key, which would have left the real database unreadable. It now keeps the key.
• Updated components. The editor, and the components that clean pasted content, read PDFs and parse web pages, are on their latest fixes.

A New Icon

A plume tucked under a notebook's band. The menu bar icon follows it and adapts to light and dark menu bars. On macOS 26 and later, the app icon follows the system's icon styles.

Getting Ready for Canto 1.0

This release changes how Canto's updates are signed so the next major version can arrive as a normal update. Nothing changes for you now.

Close this tab whenever you're ready — it won't appear again.

Local model

• Qwen 3.8 27B replaces Qwen 3.6 27B. If you were using the older 27B model, Canto switches you over automatically.
• Turbo is built in — turning on the faster engine no longer needs a second download.

Safer vault

• Apple Notes refresh keeps a copy of your current version before applying an incoming change. If it can’t, that note is skipped instead of overwritten.
• Sync does the same for changes arriving from another device.
• Backups stay in their own lane — if two vaults share a backup folder, one no longer cleans up the other’s snapshots.
• Program-like attachments open in Finder instead of running.
• Scoped chats stay private from outside apps that read your chats through Canto.

Also in this release

• Hide the scope banner in Settings → Chat if you’d rather keep the transcript clean. The scope still applies.
• If a scope points at nothing (for example a folder you deleted), Canto says so instead of quietly searching the whole vault.

Free AI queries

• Charged when work begins. A free query is spent when the model actually starts generating, not the moment you press send.
• Nothing lost to a false start. Cancel a turn, or hit a problem before generation starts, and your query is returned.
• Counted once, everywhere. Canto Chat, Document Chat, and Vault Manager all follow the same rule, and editing a message to resend it or regenerating an answer now counts like any other turn.
• A number you can trust. Some turns in 0.10.0 finished without counting at all, so the remaining count could drift. It now reflects what you've really used.

Fixes

• Truthful AI status. With the local model turned off and a chat pointed at an external or cloud model, the sidebar read "AI Off" even while the chat was working. It now shows External AI Ready, Cloud AI Ready, or Generating… along with the model's name.
• Clearer publish preview. The summary of what's excluded now states plainly that private notes, chats, and local files are removed.

🌐 Publish a note to the web

• Unlisted by default — not indexed by search engines; only people you send the link to will find it
• Looks like your note — images, math, diagrams, charts, references, and the evidence ledger all come across
• You see what's included — Canto summarizes exactly what's going out, and what's being left behind, before you publish
• Snapshots, not live editing — later edits stay private until you choose Update published version
• Unpublish anytime — one click takes the page down
• Publish from anywhere — the note's status bar, the command palette (Publish note to web…), or Settings → Publishing
• Manage what's live — Settings → Publishing lists everything you've published with view counts and one-click unpublish

🍎 Apple Notes import (macOS)

• You choose the scope — browse and search your accounts and folders, tick only what you want
• Preview before you commit — see how many notes, folders, locked notes, and attachments are involved before a single note is created
• Formatting survives — rich text, lists, tables, and images arrive as real Canto content, with attachments if you want them
• Refresh later — pull in changes you've made in Apple Notes since the import without redoing everything
• Conflicts are reviewed, never overwritten — if you edited a note in both places, Canto keeps your version and hands you a review list
• Resumable — a long import that gets interrupted picks up where it stopped

🧠 "Remember this" — memory you approve

• Capture in one click — hover any chat message and choose Remember; save a decision, commitment, preference, project fact, open question, or correction
• Scoped the way you work — keep something personal, or attach it to a single Project so it only applies there
• See when it's used — remembered items appear in the chat's context receipt, so you know what shaped an answer
• Easy to undo — open the source, forget it, mark it resolved, or replace it with a newer version
• One place to review — the Continuity Ledger tab lists everything Canto knows, filterable by scope and type
• Optional suggestions — off by default; when enabled, Canto proposes things worth remembering after a conversation using only your on-device model, and every suggestion waits for you to keep, edit, or dismiss it

🗑️ Cleanup you can trust

• Delete folders and files together — select a mix in the sidebar and remove them in one action
• Exact counts first — the confirmation shows precisely how many folders, notes, notebooks, and canvases are involved, including everything nested inside
• All or nothing — the delete either completes fully or leaves your vault untouched; no half-finished cleanups
• Catches late changes — if the folder tree changes while the dialog is open, Canto asks you to re-review the updated numbers
• Smarter Trash restore — restoring an item whose original folder is gone now puts it in the nearest surviving folder

🔋 The free AI budget now refills

• 25 queries to start, then 5 more every week — free, forever
• Honest countdowns — budget pills read "refills in 2d" instead of looking like a dead end
• Consistent everywhere — Welcome, Canto Chat, and Document Chat all show the same remaining count and refill time

🛠️ Improvements and fixes

• Connected tools stay connected — Canto's link for outside apps now heals itself after sleep, unlock, or a temporary network hiccup
• Honest connection status — see whether it's starting, running, or retrying, with a Restart now button when you want to force it
• Corrected setup snippets — copy-paste configurations for popular external tools were refreshed to match their current formats
• Sturdier Settings — if one settings panel runs into trouble, the rest of the window keeps working instead of going blank
• Faster settings navigation — jump straight to Import & Connections from the command palette
• Cleaner Projects setup — the project editor's folder picker is more compact and searchable, and closes with the keyboard
• Lighter on memory — fixed a leak in the background memory monitor
• Steadier editor — improved stability when switching between notes
• Safer folder moves — moving items between folders is handled more reliably