Every release, newest first. Betas are labelled; a beta install follows betas, a stable one stays stable. Canto and Tulip also show What’s New after each update.
May 132026
Canto 0.7.6
May 13, 2026
stable
🪟 Sidebar polish
The sidebar "island" picks up its own background, border, and hover colors — child panels like Related Notes inherit the lifted look instead of clashing with hardcoded grays.
Elevation and tinting adapt to both light and dark palettes.
Read all the notes
👋 Welcome screen
Recent note rows use your theme's text, surface, and border colors instead of fixed grays.
The New Note / New Notebook / Open Vault tiles glow with your accent color on hover.
Keyboard shortcut hints sit on the same elevated surface as the rest of the panel, in any palette you pick.
🅰️ Editor & chat typography
Shared headings, lists, and blockquotes between notes and the agent panel.
Calmer code blocks — inline code, fenced blocks, and tables get matching treatment in notes and chat, with a refined look in dark mode.
Smoother streaming entrance — assistant responses slide in with a soft animation that fully respects macOS Reduce Motion.
📂 File tree
Right-aligned metadata column for folder counts and note timestamps.
Pin and delete hover actions clustered together on the same edge.
Dense mode can now show a touch of preview metadata where it makes sense.
Refined Pinned and Private section headings with quieter labels and counts.
🧮 Currency vs. math
Smarter detection so $…$ is only treated as a math equation when it actually looks like LaTeX.
The same guard now applies in the editor, in agent chat, in notebook markdown cells, and in note & notebook history previews.
Real equations still render as before.
🤝 Agent approvals
One approval per turn — once you say yes to remote content, the agent won't ask again for the rest of that conversation turn.
The activity rail clearly shows read tools as in-progress instead of looking idle.
🐛 Fixes
Opening the What's New release notes tab no longer triggers Memory Links errors in the background while the embedding model is still loading.
Memory Links and Backlinks panels now stay quiet on non-note tabs (release notes, the graph, the vault manager) instead of trying to find related notes for them.
The embedding model now properly waits for an in-progress load to finish, so features that depend on it never run too early on launch.
Watch the agent write, live. When the agent rewrites part of a note, the new text now streams into the chat in real time so you can read along instead of waiting for a wall of text to appear at the end.
A friendlier activity timeline. Searches, reads, previews, and renames in the agent panel now read like plain English — e.g. "Renamed Old Title → New Title" and "Searched your vault for weekly review" with the matching notes listed below.
Sidebar settings tab. A dedicated Sidebar tab in Settings lets you adjust list density and how the Private list behaves.
Memory, simplified. The Memory window now offers one clear choice: Adaptive (Canto picks the best context size for your Mac) or Manual (you set it). The old "Higher Context Mode" toggle is gone.
Read all the notes
💅 Improved
Smoother streaming. Incoming text gently fades in character-by-character with a small catch-up so fast bursts still feel responsive. Fully respects macOS Reduce Motion.
Calmer panel motion. New chat rows and timeline entries glide in with a soft entrance instead of snapping.
Auto-scroll keeps up. Long live previews keep the conversation pinned to the bottom without yanking you around mid-read.
Multi-step edits, clearer. When the agent applies several edits in one turn, each one shows up as its own card with a clear drafting → applying → done state, indented neatly under the parent step.
Faster link previews. Canto can now fetch up to four link previews at once on a dedicated background thread, so a clipboard full of URLs no longer slows the app down.
Lighter background work. Background indexing now touches only the links that actually need a preview instead of re-scanning every note, and briefly steps aside while you're typing so the editor stays snappy.
Honors your link auto-fetch setting. If you've turned auto-fetch off, the background catch-up loop now stays off too.
Pre-selected new-note title. Open a fresh note and the placeholder title is already highlighted — just start typing.
Esc closes history modals. Both the note and notebook version-history windows now dismiss with Esc, like the rest of the app.
Clearer Memory copy. The context-size tooltip and helper text explain each mode in plain language.
🐛 Fixed
No more false "model stuck" warnings. When an agent step is waiting on you (an approval or review), Canto no longer treats that pause as the model being frozen.
Sturdier tag sync. When two devices add the same tag at the same time, Canto now merges them cleanly instead of leaving duplicates behind.
Renames into empty notes. Renaming a note whose body starts blank now correctly syncs the new title into the first line.
Drag Markdown files into Canto. Drop one or many .md files onto the sidebar and Canto turns them into real notes — into Overview, into your Vault, or directly inside a folder. Multiple files are imported in one pass, and a highlight shows you exactly where the note will land before you let go.
More appearance themes. The Appearance settings now include a curated set of popular light and dark looks beyond the original Zen, Mist, Midnight, and Monokai. Every preset is properly aware of light/dark mode, and switching applies instantly across the sidebar, editor, and agent panel.
Read all the notes
🔗 Links
Better link previews. When a site blocks previews or strips its title, Canto still gives you a clean, recognizable link pill instead of a broken one.
Edit link pills inline. Change a link's display text or URL right where it sits in the note — no separate modal dance.
Quieter retries. Flaky network responses are retried in the background, so previews fill in once the page actually answers.
Tidier link insertion. Adding a link from the modal feels a touch more polished and predictable.
🔐 Sync
More reliable auto-unlock. Opening Canto on a second Mac is steadier, even when the cloud side is still catching up.
Cleaner reconciliation. When devices need to compare notes, Canto handles the back-and-forth with fewer false alarms.
Better checkpoint recovery. If a sync session is interrupted, picking up where it left off is more graceful.
🧹 Attachments
Cancel actually cancels. Stopping an attachment's indexing now truly halts the in-flight work instead of letting it finish quietly in the background.
Honest progress. The embedding queue inspector reflects the real state of what's running, paused, or done.
Lighter on your Mac. Cancelling frees CPU and memory right away — most noticeable when you're juggling big PDFs or videos.
🎨 Polish
Sidebar settings, repositioned. The gear in the AI island is back at a more reachable spot for thumb-and-trackpad users.
🛠 Under the Hood
Added safety checks that protect link fetching and sync device handling, so the things that work today keep working tomorrow.
Old attachments now clean themselves up. Once every linked Mac agrees an attachment is no longer in use, Canto safely removes it from your synced vault so iCloud storage stops growing forever.
One-hour safety window. Orphaned attachment files have to age for at least an hour before they're eligible for cleanup, so freshly removed files are never at risk.
Empty folders get pruned too. After cleanup, the now-empty storage folders are removed as well — no more thousands of empty placeholders in iCloud.
New Blob storage view in Sync Settings. See exactly how many attachments are protected, eligible for cleanup, or being held back, with a per-file breakdown if you want it.
Read all the notes
☁️ Friendlier with iCloud
Patient with lazy downloads. When iCloud is still pulling a file in the background, Canto now waits and quietly retries instead of reporting a sync failure.
Fewer scary error messages. Transient "missing payload" warnings during normal iCloud catch-up no longer surface as failures.
Self-healing attachments. If your local attachment index ever goes missing, Canto rebuilds it from your retained sync history so files aren't accidentally dropped.
✍️ A more disciplined writing agent
Section-sized drafts. Long notes are now written one section at a time in deliberate passes, so big drafts feel less jarring as they fill in.
No more duplicated titles. If the AI tries to use your note's title as the first body heading, Canto blocks it and routes the agent to rename the note properly instead.
Better mid-draft repairs. When the agent needs to fix the end of a paragraph it just wrote, it repairs the tail instead of appending a near-duplicate copy.
Safer handling of equations and special blocks. Fragile inline elements are only dropped when explicitly necessary, not as a side effect of a normal edit.
📊 Honest storage numbers
Counted from disk. Sync Settings now reports journal counts based on what's actually in your vault folder, with the database count shown alongside if the two ever drift.
Quieter idle behavior. Canto no longer rolls over its sync history files just because the app sat idle — the safety backstop now waits a full week.
iCloud bytes at a glance. The compact status line in Sync Settings now shows how much space your attachments are using inside the synced vault.
🛠️ Smaller fixes & refinements
Cleaner Repair Full Sync failures. If a repair hits an error before it can publish, Canto stops early and tells you what happened instead of writing a half-finished state.
Better mid-segment catch-up. Devices returning to sync in the middle of a busy edit window now pick up the remaining changes correctly on the first try.
Fresher attachment index. Canto refreshes its attachment index roughly every five minutes while unlocked, so storage cleanup always works from up-to-date information.
Web pages are evidence, not instructions. Fetched pages, PDFs, YouTube transcripts, and search snippets are now clearly labeled as untrusted external content, so a sneaky webpage can't impersonate your prompt or pretend to be a system instruction.
Hidden instructions in webpages are ignored by design. If a page tries to tell the AI to delete notes, leak information, or run other tools, the agent treats it as quotable text — not as a command.
Same useful workflows. "Summarize this article into my note," "compare these two docs pages," and link-pill citations all keep working exactly as before.
Read all the notes
✅ A clear approval step after reading the web
Edits pause for approval after a URL or web search. If the agent fetched a link, ran a web search, or pulled in content from a saved link during a turn, the next note edit in that turn asks for your approval — even if you normally have auto-approve on.
You see why approval is being asked. A small note explains: "Remote URL/search content was read in this run, so edits require approval."
Read-only summaries stay friction-free. If the agent only quotes or summarizes a page back to you in chat, no extra approval is needed.
⚙️ New setting: Auto-approve after approved URL content
Found in Settings → Agent.
On (default): once you approve the first URL-grounded edit in a run, Canto returns to its normal auto-approve behavior for the rest of that run.
Off: every URL-grounded edit asks for its own approval, for the most cautious workflow.
Per-run only. A brand-new chat starts fresh either way, so trust never carries over silently between runs.
🔒 Safer URL fetching
Public web only. Links that try to reach your machine, your local network, or other private/internal addresses are rejected before any request goes out.
Validated redirects. Every hop in a redirect chain is checked against the same rules, so a public URL can't quietly bounce to a private one.
Stricter URL rules. Only normal http and https URLs are fetched, and links with embedded credentials are turned away.
Same surfaces, safer plumbing. Applies to Agent Chat URL prefetch, the Fetch this page tool, and your saved link pills alike.
🗂️ One approval flow for all vault changes
Renaming, moving, deleting, duplicating, and pinning notes, plus creating and renaming folders, now consistently route through the same approval UI as note edits — so vault-changing actions can't slip through as if they were read-only.
🛠️ Smaller refinements
Tighter agent prompt rules. The agent is now explicitly told to ignore instructions found inside fetched pages or search results.
Clearer grounded steps. When the agent uses web evidence, it's easier to see at a glance which steps were grounded in remote content.
Quieter, more predictable runs. Approval prompts only appear when something genuinely warrants confirmation.
Agent reads URLs you paste. Drop any link into Agent Chat and Canto fetches the page in the background while you finish typing — by the time you hit send, the AI can quote, summarize, or compare its real contents instead of guessing from the URL.
"Fetch this page" tool. The agent can also pull in new URLs mid-conversation when it decides it needs them.
Unified agent timeline. Tool activity in Agent Chat is now one continuous vertical rail with status dots and a connecting line, instead of a stack of separate cards.
Smart activity grouping. Long runs collapse into a single line like "Reviewed note · 6 actions · 2.1s", and whatever step is currently running pins itself open so you can watch progress without scrolling.
Glow where the agent edits. Lines the AI inserts, deletes, or rewrites briefly glow in your accent color and then fade — so it's obvious what just changed, with nothing left behind to clean up.
Adaptive Memory Mode by default. Canto now picks sensible memory caps based on your Mac's RAM: more headroom on 32 GB / 64 GB machines, a gentler footprint on 16 GB ones — no setting to flip.
Read all the notes
🚀 Faster
Quicker startup. The sidebar, recents, and search no longer load full note bodies just to draw a row.
Lazy tab restore. Open tabs are brought back on demand instead of all at once at launch.
Faster Memory Links on big vaults. Note similarity uses an approximate nearest-neighbor index for a clear speedup on large note collections.
Smoother background work. MiniLM (the model behind Memory Links and search) runs in tighter batches and prioritizes the note you're currently viewing.
Snappier local agent replies. Canto picks the right tools for each turn and reuses the AI session between turns when it can.
Lighter memory pressure. Canto only trims its caches when memory is genuinely tight, instead of jumping at every blip.
🎙️ Audio & Video
Native Whisper transcription. Audio attachments and videos now transcribe with Canto's built-in Whisper engine end-to-end — faster, more reliable, and no extra setup or helper download.
Quiet retries. Older audio or video items that failed under the previous helper are picked back up automatically.
🎨 Polish
Clearer Sync & Vision status chips. The small status indicators now share one consistent style with calmer colors and more readable labels.
Bigger sidebar AI actions. The action buttons inside the sidebar's AI island are easier to hit and easier to scan.
More reliable window dragging. The top of the sidebar is a sturdier drag region, so moving the window won't accidentally toggle controls.
One-surface shell. The editor, sidebar, and status bar share a single calm background tone, so the app reads as one piece.
Tighter sidebar typography. Another pass over labels and spacing for a quieter, more confident look.
🛠️ Fixed
Sturdier agent edits. Long AI edits are less likely to drop content thanks to a stricter safety check on what the agent can write.
Full note-agent toolkit restored. Every note-side automation tool is available again, with bounded previews so the agent doesn't choke on huge notes.
Quieter sync. Duplicate journal records no longer cause noisy retries on iCloud Drive.
Smaller agent payloads. Sidebar previews sent to the AI are trimmed so it gets the gist of nearby notes without the full text.
Canto
Tulip
opsty
subtank