Changelog.

Agent

• Added a new Agent tab in Settings where you can enable auto-approve mode — when on, the agent applies edits to your notes directly without stopping to ask for confirmation at each step
• A warning banner in the Agent tab reminds you that auto-approve skips the diff review, so you always know the mode is active before turning it on

AI Models

• The AI Models settings page is cleaner — the Low-Memory Loading and Higher Context Mode options now sit side by side instead of stacked, and all the labels are shorter and easier to read at a glance

Ollama

• When you have an Ollama endpoint connected, Canto now reads the actual context window size from whichever model is currently running, so prompts use the real available context instead of a generic fallback

Attachments

• Fixed a bug where attachments added by the agent, or restored after an undo, were sometimes not indexed for AI search
• Fixed a bug where undoing and redoing a change could accidentally delete the search index for an attachment that was still in your note
• Attachments now survive brief disappearances (such as when the agent rewrites surrounding content) without losing their indexed data

Improved Claude Code setup

Connecting Canto to Claude Code is now much more reliable.

• Fixed a problem that could prevent Claude Code from connecting to Canto's MCP server
• Bearer-token based connections now behave correctly instead of failing in a confusing way
• The setup flow is smoother when using Canto with external AI tools

Easier re-registration

If Claude Code or another tool gets stuck with old MCP credentials, Canto now makes recovery much easier.

• Added a ready-to-copy remove-and-re-add command sequence for Claude Code
• Makes it much faster to refresh credentials without guessing which commands to run
• Helps recover from stale or cached MCP setup issues

MCP settings panel is fixed

The MCP Server settings page is now easier to use when authentication is enabled.

• Fixed a layout bug where turning on a bearer token could break the right side of the settings UI
• Copy buttons and controls stay accessible
• Long configuration snippets now stay contained in the panel instead of pushing the layout out of place

In short

0.5.3 makes the MCP Server feature feel much more solid: easier to connect, easier to recover if something goes wrong, and easier to use in Settings.

New Features

• Audio file transcription — Drop an MP3, WAV, M4A, OGG, FLAC, or AAC file into any note and Canto automatically transcribes it in the background. No API key or internet connection required — everything runs on your device.

• Timestamped transcript in the note — Expand an audio attachment to see the full transcript below the player. Each line is paired with a timestamp, and clicking any timestamp instantly seeks the audio to that moment.

• Copy transcript — A single button in the transcript header copies the entire transcript text to your clipboard.

• Audio is searchable by AI — Once transcribed, the audio content is indexed just like your notes and documents. You can ask your AI assistant questions about what was said in a recording.

• Live Transcription panel — A new floating panel (⌘P → "Live Transcription") lets you record your microphone and watch the transcript appear in real time. When you're done, insert the result directly into the note you're working in or copy it to the clipboard.

• Mic level meter — The Live Transcription panel shows a live audio level bar so you can confirm your microphone is being picked up before you start talking.

Read all the notes

Improvements

• Transcription progress on the attachment pill — While an audio file is being transcribed, the attachment pill shows a live percentage so you always know how far along it is.

• Transcript can be cancelled — Right-clicking an audio attachment while it's transcribing gives you a "Stop Indexing" option, the same as for images and videos.

• More audio file types recognised — Files with .m4a, .ogg, .flac, .aac, and .oga extensions are now correctly identified when dragged from Finder, even when macOS doesn't supply a file type automatically.

• Transcription survives restarts — Audio files queued for transcription are tracked in the same persistent queue as documents and images, so if you close Canto before transcription finishes, it picks up where it left off on the next launch.

Canto 0.5.1

Indexing Picks Up Where It Left Off

If Canto closed before finishing the indexing of a PDF, image, or video, that work was lost. Now the queue is saved to disk — close the app mid-way through, reopen it, and indexing continues automatically.

Dropped Attachments Queue Automatically

Every file you drop is now registered the moment it lands in a note. If the AI model needed to index it isn't ready yet, the attachment waits in line and is processed as soon as the model becomes available — no manual retry needed.

Drop the Same File Twice, Index It Once

If you drop the same PDF, image, or video into multiple notes — or into the same note again — Canto recognizes it and reuses the results from the first time. No re-reading, no re-processing, instant.

See Whether an Attachment Is Indexed

Each PDF, image, and video pill now shows a small colored dot:

• 🟢 Green — indexed and searchable by the AI
• 🔴 Red — not yet indexed

Right-Click to Control Indexing

Right-clicking any PDF, image, or video now gives you direct options:

• Index Now — jump to the front of the queue and start immediately
• Add to Index Queue — queue it normally
• Stop Indexing — cancel an in-progress job
• Remove Embeddings — clear the stored index for that file

Embedding Queue Inspector

A new panel shows everything currently in the indexing queue. Click the layers icon in the status bar (it appears when files are pending) to open it. You can see what's waiting, what's actively running, what failed, and cancel individual entries.

Delete Vision Models to Free Space

You can now delete downloaded vision model files directly from Settings → Image & Video. Useful if you downloaded a model to try it and want to reclaim the disk space afterward.

Vision Server Status in the Status Bar

A small indicator at the bottom of the editor shows whether the vision server is running or idle, so you always know its state without opening Settings.

Bug Fixes & Improvements

• Fixed: Dropping a file before a vision model was downloaded would silently skip indexing with no feedback. Now a notification appears with a link to download a model.
• Fixed: A stale vision server process left over from a previous session could prevent the new one from starting. Canto now cleans it up automatically.
• Fixed: Deactivating a license and reactivating in the same session could leave the embedding model in a broken state. This is now reset correctly.
• Improvement: The audio transcription model (Whisper) is warmed up in the background at startup, so the first video you drop starts transcribing right away instead of waiting.

Ask Your AI About Documents

• Drop a PDF, Word doc, or text file into a note and Canto will index it automatically
• Ask the AI questions about any indexed document and get answers with page-level citations
• Citation style is configurable — choose IEEE, APA, or footnote format in Settings → Notes
• Hover over a citation badge in the chat to preview the source passage and page number
• Document embeddings are included in workspace exports and scheduled backups

Ask Your AI About Images

• Images dropped into notes are now described by an on-device vision model
• The description is embedded and searchable through Memory Links and the Knowledge Graph
• Three vision models to choose from: Qwen3 VL 4B (recommended), Qwen3 VL 8B, or Gemma 4 E2B
• Download your preferred model once from Settings → Image & Video

Ask Your AI About Videos

• Video files are analyzed scene-by-scene using keyframe extraction and the vision model
• Audio is transcribed with Whisper, producing timestamped transcript segments
• Ask questions like "what was shown around the 2-minute mark" and get precise, time-aware answers
• Processing happens entirely in the background — the editor stays fully responsive

Performance Improvements

Memory Links Loads Instantly

• MiniLM (the embedding model behind Memory Links) is now bundled with the app
• Load time dropped from 3–15 seconds to under a second on every launch
• No internet connection needed for Memory Links to initialize

Whisper Transcription Starts Immediately

• Whisper ONNX model files are now bundled with the app
• First-use transcription cold-start dropped from ~2.5 minutes to under 10 seconds
• Whisper runs in a background thread so the UI never freezes during transcription

Fully Offline from First Launch

• Memory Links, document indexing, image understanding, and voice dictation no longer require a network connection on first use

Read all the notes

Stability Improvements

Safe Multi-Attachment Processing

• Dropping several files at once no longer risks crashing the app
• All embedding jobs are now serialized — one runs at a time, the rest wait their turn
• Attachment pills show a Queued badge while waiting, and a labeled progress badge while active (e.g. Describing 33%, Transcribing 67%, Saving)

Attachment Cleanup on Delete

• Removing an attachment from a note now also deletes its stored embeddings
• No more stale data accumulating in the database over time

File Drop Detection Fixed

• Files dragged from Finder that don't expose a MIME type (common on macOS) now correctly detect their type from the file extension, so indexing triggers as expected

Settings & UI Changes

Data Directory — Storage Stats

• The Data Directory settings page now shows your encrypted database size on disk
• Indexed content counts (images, video scenes, transcript segments, document chunks) have moved here from the Image & Video page, where they now sit alongside storage information

Calendar Starts on Sunday

• The daily notes calendar view now lays out weeks starting on Sunday

📅 Calendar View for Daily Notes

Say goodbye to the "last 7 shown" limit. Your daily notes now have their own Calendar tab in the sidebar, giving you a full month-at-a-glance view of where you've written and where you haven't.

• Browse your entire history — navigate forward and backward through months to find any day you've written
• Tap any date to open an existing daily note, or create a new one if that day is empty
• Right-click a date for quick actions like opening or deleting a note
• Visual at a glance — a small dot under each date shows you which days have notes, so you can spot writing gaps instantly
• Timezone-smart — dots appear on the correct days based on your timezone settings

Your Quick "Today" shortcut in the Overview tab is still there for convenience — the calendar just gives you the full picture alongside it.

🧠 Smarter Agent Sessions When You Switch Tabs

When you ask Canto's agent to help with edits and then switch to another tab while it's working, it now stays on track without losing its place.

• No more lost context — the agent keeps working in the right document even if you clicked away
• No more frozen approvals — if the agent asks for your OK before making a change, that prompt won't disappear when you switch tabs
• Accurate changes in the right spot — the agent reads and writes to the correct editor even after a mid-task tab switch

📓 Better Notebook Editing for Agent Automation

When agents use Canto's automation tools to work with notebooks (like code/test notebooks with cells), they now work with numbered cells instead of ambiguous references.

• Clearer targeting — cells are numbered 1, 2, 3... with no confusion about which is which
• Safer edits — invalid cell references are caught upfront instead of accidentally modifying the wrong cell
• More capable agents — agents can now open notebooks and read cell contents through automation, letting them inspect structure before making changes
• Readable logs — tool summaries show clean, human-friendly cell numbers

---

Read all the notes

Stability & Behind-the-Scenes

• Refined the daily note creation flow to support creating notes for any date, past or future
• Behind-the-scenes cleanup to make the entire agent editing system more consistent and dependable

Canto 0.4.2 makes AI editing feel more reliable and more focused. This release improves how edits are placed inside notes, handles structured content more carefully, and removes a few older experimental writing features to simplify the overall experience.

Highlights

Smarter AI note editing

AI edits are now better at understanding where a change should happen inside a note.

• Better placement when you ask for something to be added below a section or under a heading
• More accurate targeting for tables, lists, checklists, code blocks, math blocks, and Mermaid diagrams
• More dependable edits when working around existing sections instead of only writing at the top or bottom

Better handling for structured content

Notes with more complex formatting now behave better during AI edits.

• Safer edits for lists and checklists, including multi-step changes
• Better handling of tables so row and column requests are less likely to be treated like plain text
• Clearer editing behavior for code, math, and diagram blocks
• More precise section-based inserts for summaries and follow-up content

Simpler AI writing experience

A few older experimental AI writing features have been removed to make the core editing flow cleaner and more consistent.

• Removed inline ghost-text suggestions
• Retired tab-style autocomplete writing
• Removed writing style profile options from Memory Links settings
• Removed Continue Writing shortcuts in favor of a more consistent AI editing flow

A more focused Memory Links experience

With the older autocomplete features gone, Memory Links is now more clearly centered on note connections.

• Cleaner settings with fewer confusing options
• Stronger focus on related notes, embeddings, and the Knowledge Graph
• Updated onboarding copy to better reflect how the feature works today

Model naming cleanup

Local Qwen 3.5 model names were cleaned up internally to be more consistent.

• Existing saved model selections are carried forward automatically

Stability and polish

This release also includes behind-the-scenes cleanup to make AI behavior more consistent and dependable going forward.

This update improves the new Vault Manager experience introduced after 0.4.0, making it more capable, cleaner to use, and easier to trust when applying changes.

What's improved

• Vault Chat can now handle more organization tasks

  • In addition to planning moves and tag updates, it can now help with more note and folder cleanup actions such as renaming folders, deleting empty folders, duplicating notes, pinning notes, and removing notes when appropriate.

• Better visibility while the AI is working

  • Vault Chat now gives clearer live feedback while it scans your vault and prepares changes, so it feels less opaque and more guided.

• Improved draft plan review

  • Proposed changes are easier to scan before applying, with clearer grouping and visual cues for different action types.

• Smarter behavior based on your selected AI model

  • Vault Chat is now better aligned with the model you're using, helping responses and planning feel more consistent.

Read all the notes

Design polish

• Cleaner, less boxed Vault Manager layout

  • The Vault Manager view has been simplified to feel more modern and less cluttered, with a flatter layout, fewer heavy containers, and better use of space.

• More streamlined vault overview

  • The stats, folder tree, and tag view are now presented in a cleaner way, so the left side feels more like a natural workspace and less like a dashboard full of cards.

Fixes

• Failed changes now explain what went wrong

  • If part of an approved plan fails, the chat now shows the actual reason instead of only saying that something failed.

• Error messages are easier to understand

  • Failure details are now written in more human-friendly language rather than internal action names.

Under-the-hood improvements

• Compatibility and stability updates
  • This release also includes dependency and toolchain refreshes to keep the app stable and current.

If you liked the Vault Chat concept in 0.4.0, 0.4.1 makes it feel much more complete: more actions, clearer feedback, cleaner presentation, and better error reporting.