Changelog

Added

• Peek into another workspace without leaving this one - ask how you handled something in a different workspace and Mira answers from it directly, read-only. You stay where you are.
• Ask Mira to switch workspaces - say the word and it moves you, waiting until the reply has finished so nothing gets cut off mid-thought.
• Background processes you can see and stop - dev servers and long-running commands now appear in Background Tasks with their address and a Stop button that actually stops them. Servers you launched yourself show up alongside Mira's.
• Progress updates worth reading - before a long stretch of work Mira says what it is about to do, then speaks up again only when the approach changes, something surprising turns up, or it hits a wall. Routine steps stay quiet, because the activity list already shows them.
• Plan runs you can watch - a running task shows live presence, keeps its controls while it runs, records where the run came from, and reports real timing and token counts when it lands.

Changed

• A conversation you can actually read - the chat is now a single centered column with a comfortable line length, instead of prose stretched across the full width of the panel. Conversations also moved out of horizontal tabs into a rail shared with the pop-out Mira window, so a long list stays navigable instead of squeezing every tab down to a sliver.
• A rebuilt activity view - every step reads as one line with its result attached ("Grepped App.css: 41 matches in 12 files"). Runs of routine lookups fold into a single receipt that keeps their findings, reasoning collapses to a one-line headline you can expand, and each row carries a small icon for the kind of work it was. Changes and failures never fold away, and a live timer runs while work is in flight.
• Code and tables look right in a reply - fenced code now gets syntax highlighting with a language label and a copy button, in both light and dark. Wide tables scroll on their own instead of blowing out the column.
• Messages queue instead of interrupting - typing while Mira is working parks your message as an editable chip. Reorder them, edit one in place, send one immediately, or use one to steer the turn already running.
• Mira remembers what the work produced - memory used to record only what you stated outright. It now also keeps what a job actually established.
• The folder journal is a visible UNDRA.md - Mira's own working notes inside a connected folder used to sit in a hidden dotfile nobody ever found. They now live in UNDRA.md at the folder root. Your house rules stay in AGENTS.md: different files, different owners, and Mira will not rewrite yours. Existing journals keep working and move across on the next update.
• Mira reads the rules on the door first - working in a connected folder it has not touched in that conversation, it now gets that folder's AGENTS.md and UNDRA.md before its first action there rather than after.
• Search says when it cannot answer - an unbuilt or unavailable search index used to come back as a confident "nothing found". It now tells you the index is still building, and answers from what it can read directly in the meantime.

Fixed

• Work you just did can no longer go missing from search - a note written a minute ago was invisible until the index caught up, so Mira could report something you had just recorded as not existing.
• Search inside a folder-scoped chat returns results - it was coming back empty every time, whatever your index actually held.
• Large connected folders list quickly - listing a big repository kept crawling the entire tree, node_modules included, long after it had enough results.
• A dev server that has stopped stops looking alive - its preview offers a restart instead of pretending to still be running.
• A process you stopped reads as stopped, not failed - including when you close it from the pop-out Mira window.
• Moving an item to another folder sticks - a leftover ordering file could quietly pull it back.
• Development-only tools can never appear in a released build - they are left out of the toolbox entirely rather than merely hidden.
• The Background Tasks panel scrolls again, and the workspace window controls sit in the true corner.

Added

• Run any task in a plan - every open task is runnable as it stands: the title and description you wrote are the instruction, and clicking Run is the only opt-in. The separate agent-brief field is gone - if a task is too vague to act on, the run comes back asking for what it needs instead of guessing.
• Every plan run is saved as a transcript - a run used to be a black box that left a single summary line behind. Now each one is kept as a full conversation (every step, tool, and command output) and a View run button on the task opens it.
• Browser-style navigation between notes - clicking a link to another note now reuses one preview tab instead of piling up a permanent tab per hop, while Ctrl+click still opens a permanent one. Your mouse's back and forward buttons walk the trail, as do Alt+Left and Alt+Right, and hovering a link peeks at what's behind it without opening anything.
• Turn a conversation into a skill - type /skill in a chat and Mira distills what you just worked through into a reusable skill note, then opens it for review. It takes effect immediately, with one-click undo if it isn't what you wanted.
• Your skills are slash commands - any skill note you've written now appears in the / menu under its own name, so you can invoke it directly instead of describing it.
• One-click Update Claude Code - if your installed Claude tool is out of date, newer models simply don't appear in the model picker, with nothing to explain why. Undra now spots that gap and offers an update button right where you'd notice it. When you're current, there's nothing to see.
• Undra names the app holding a folder - deleting a connected folder that another program has open now opens a dialog naming the program, and finishes the delete by itself the moment you close it, instead of a message telling you to guess and try again.
• Stop all dev servers - one command to shut down every dev server Undra is running.

Changed

• See what every AI provider is doing - Codex used to work invisibly: the transcript jumped from "Thought for 31s" straight to the answer, with every search, file read, and command hidden. All of it now shows up as normal activity steps. Cursor turns shed their internal housekeeping rows and raw output, and commands read as plain text on every provider instead of a wall of escaped path characters.
• Thinking reads as prose again - the assistant's reasoning was being chopped into a new line every few words. It now flows as paragraphs.
• One Storage & Cleanup card - Settings, Workspace merges the old Storage and Media Maintenance sections into a single check, review, confirm flow in place of five separate buttons. Cleanup now refuses to run at all if it couldn't read every document's media references, rather than risk deleting something still in use.
• Extension switches are honest and safe - turning a bundled extension off can no longer leave the app unable to start, switches that can't be turned off now say so up front with the real reason, and turning off Portal Preview correctly hides its right-click entries.

Removed

• Item slash commands - /new-note, /rename-item and the rest are gone; the explorer does those faster and Mira does them better. The / menu is now session controls and power workflows only.
• Per-task agent briefs - replaced by running the task text itself. Plans you already have keep working.
• AI Prompt from the canvas right-click menu.

Fixed

• Videos work properly - a video showed a black thumbnail in your media library, played as a black rectangle on a canvas, and couldn't be scrubbed. Undra was serving video files in a way that left them unseekable, so every seek quietly snapped back to the start and thumbnails captured a black opening frame. Videos now seek, scrub, and get a real thumbnail chosen to avoid a black frame.
• Plan tasks no longer lose their contents during an agent run - a background refresh could push an emptied version of an open plan into the editor, and your next save made that loss permanent.
• Asking Mira to run a plan by name runs the right one - a plan named exactly what you typed could lose out to an unrelated note, so the wrong plan ran. Mira now also says it ran a plan rather than edited an item, and tells you what to do when every task is already finished instead of dead-ending.
• A slash command typed mid-reply no longer reaches the model as a message - typing /clear while Mira was answering sent her the literal text.
• View run works when the chat panel is closed - it opens the panel for you.
• Open in image editor shows your image instead of blank canvas chrome.
• Mira sees your connected folders from the pop-out window - folder tools reported none at all there, which dead-ended everything that followed.
• Emptying the trash works for folders containing code repositories - read-only files inside them made the delete fail.
• Renaming a workspace in Home shows the selected text clearly, so replacing the name is obvious.
• Connected folder renames are atomic, and Undra no longer holds a lock on managed folders on Windows.

Changed

• Menus stay above live previews - a running preview (a project's page or a connected folder shown inside Undra) paints over the app, so menus that opened on top of it were getting hidden. Now the tab menu, right-click menus on the sidebar and tabs, the split-view drop indicator, the tab-stack popover, and the breadcrumb folder picker all appear above the preview where you can actually see and use them.
• Tidier connected-folder previews - opening a connected folder's page now gives you a single preview tab titled with the folder's own name, and stacked previews keep their names instead of all reading "Preview".
• Questions you answer stay in the chat - when Mira asks you to pick between options and you choose, the answered card now stays in the conversation as a record of what you decided, instead of disappearing as the chat moves on.
• A hardening pass - a round of internal fixes closing several classes of bug found in a security and correctness audit, for a steadier, safer app.

Fixed

• The on-device search model downloads correctly again, after the service that hosts it moved to new download servers.
• Saving into a connected folder for the first time no longer fails when its backup location doesn't exist yet, and a write that does fail now points you at how to recover it.
• The New Tab no longer flashes an unfinished background while it opens.
• The breadcrumb dropdown no longer lists connected folders, which it could never open from there anyway.

Added

• Drive Undra from your other AI tools - Undra now offers its own connection point so tools like Claude Code, Cursor, or Codex can read and change your workspace through Mira's same toolbox. Copy the connection snippet from Settings → AI → Use From Other Apps.
• Chapter markers in long chats - on a big multi-step job, Mira drops named dividers ("Research", "Implementation"), and a jump menu appears so you can skip straight to a phase.
• Side-task chips - when Mira spots something worth doing that's off your current ask, it flags a chip under the reply; one click spins it into its own new chat, leaving your current work untouched.
• A memory that tidies itself - Mira periodically cleans up what it remembers: merging duplicates, retiring stale facts, and filing loose notes under real topics, and it shows you every change for approval.
• Paste an image into your library - images you paste into chat can now be imported into your media library to keep and reuse.

Changed

• Chats keep their place when you switch away - move to another chat or tab while Mira is working and come back, and the whole conversation is still there, her live steps and streaming reply keep going, and the timers and token counts pick up where they were, instead of collapsing to a blank "thinking" spinner. This applies across the dock, chat tabs, and the pop-out Mira window.
• A more reliable Stop - the Stop button halts the turn you're looking at, including after you've switched away and back, and when several chats are running at once.
• Parallel chats stay independent - running more than one chat at a time no longer lets them interfere: sending a message in one won't interrupt another, and each keeps its own live steps, found-item cards, and "saved to memory" confirmations (so an Undo won't touch the wrong chat's memory).
• Two windows stay in sync - with the same chat open in the dock and the pop-out Mira window, they coordinate: you won't accidentally start two competing replies, and neither window overwrites the other's saved history or tabs.
• Steadier in-app previews - running a project's page inside Undra is more robust: previews no longer wedge on certain security headers, muted videos autoplay, pages keep running when their tab isn't focused, and dev servers bind to a reliable address.
• One place for AI settings - all AI configuration (providers, connected tools, memory) now lives in a single Settings surface with a clear app-wide vs. workspace split, plus a single Enable-AI switch in Undra Home.
• A newer foundation - Undra moves to a much newer app runtime under the hood, bringing a large batch of upstream security and stability fixes.

Fixed

• A chat interrupted mid-reply (an app restart or crash) now shows a "this didn't finish" note with a one-click retry, instead of looking stuck.
• A damaged chat-history file now recovers from an automatic backup instead of the conversation disappearing.
• The token and time counter under a reply no longer resets to zero when you switch chats mid-turn.
• Canvas wheel-zoom and the pen overlay track the right spot after you open or close side panels.
• Deleting a connected folder that's running a dev server gives a clear reason and points at the right program, instead of a vague "in use".

Added

• A memory you can actually read - what Mira remembers is now kept as topic notes right in your workspace (grouped by subject, with a clean index), so you can open, read, and edit any of it like a normal note instead of trusting an invisible list. Mira also suggests memories worth keeping for your approval, and lets you pin hard rules it must always follow.
• A Skills lens in the sidebar - the skills you've taught Mira get their own section above your Mira folder, so you can see what it knows at a glance. Skills now earn their keep: Mira offers to sharpen one that's getting weak and retire ones it never uses.
• Search your past chats - Mira can search across everything you've discussed before, pull up an old conversation, and pick up where it left off.
• One model picker for every provider - if you've connected more than one AI provider, a single picker in the message box lists all their models together.
• A New Tab that starts you off - the empty workspace center is now a New Tab with an inline command bar and a one-click jump into search across all your notes.
• Recover deleted folders - disconnected or deleted workspace folders now appear in Trash, where you can restore them or clear them for good.

Changed

• A refreshed look - a new default accent (a calmer steel teal in place of blue) and a consistent, hand-picked icon set across the activity bar, status bar, and every item type. Your saved accent color is left untouched.
• A faster, steadier graph - the workspace visualization stays smooth with thousands of items (a rewritten physics engine, smarter level-of-detail, lighter camera work), remembers where nodes settled between sessions so it opens laid out instead of rebuilding, and a hover card previews a node's contents without opening it. Skill notes get an amber sparkle so you can spot what Mira has learned.
• A simpler Overview - the workspace center is clean and uncluttered by default, and the dock opens chat-only, with the heavier panels there when you want them.
• Steadier long chats - older parts of a long conversation are summarized more consistently and warm chats skip redundant work, so replies stay quick deep into a session.

Removed

• New dashboards - Mira no longer creates dashboards and the create paths are gone. Dashboards you already have still open and edit as before; this just retires them as a thing you make new.
• In-app Community browse - the tab for browsing extensions inside the app is gone for now.

Fixed

• Editing files in a connected folder scrolls properly - long files show a scrollbar and the editor no longer clips.
• Deleting a workspace folder is reliable, even on network drives, with clear errors when a folder is locked by another program.
• Clicking the instructions badge on a connected folder opens its house-rules (AGENTS.md) file.
• Saves recover on their own from version conflicts on dashboards, canvases, and plans instead of erroring.
• A hung first request retries after 30 seconds (was two minutes), with a note in the footer.
• The packaged undra command-line tool loads correctly on Windows.

Added

• See what Mira remembers - AI Settings now has a Memory section that lists the long-term facts Mira has saved about you, with one-click delete (two-step confirm, since removal is permanent). Until now there was no way for you to see or clear this.
• Activity steps show where they acted - each step in the assistant's activity view now carries a small badge saying whether it touched your workspace, a connected folder, or an external location, and that scope is hoisted into the phase headers.
• Trade-offs on clarification cards - when Mira asks you to choose between options, each option now spells out its trade-off instead of just a label.
• Name a new folder as you add it - creating a workspace portal now lets you type its name inline instead of accepting a default and renaming after.

Changed

• A calmer, more consistent look - a broad polish pass across AI Settings, the main Settings, and the pop-out Agents window: softer selection and primary buttons, quieter list rows, gentler dark-mode text, and matching card styling throughout. AI Settings and Mira now sit together on the activity bar, and the pop-out Agents window is branded as Mira.
• Overview and Home are now separate - Overview (your in-workspace dashboard) is its own surface, while Undra Home stays the multi-workspace launcher. Home is a curated grid you can reorder and resize with reliable persistence, and Overview's Continue section is now a dense resume list instead of large cover cards.
• A simpler Extensions hub - the filter sidebar and animated metric strip are gone, replaced by a search box, a quiet status line, and a clean list where the always-on built-ins tuck behind a single disclosure.
• Inline Mira in notes feels alive - the ghost text Mira writes into a note now shimmers a "still writing" label while it works, and once it settles it shows a formatted preview of how the text will look, clearly marked as provisional until you press Tab to accept. Your note stays untouched until you do.
• AI edit backups moved out of your folders - snapshots Undra keeps before the assistant edits a connected folder now live in app storage instead of inside the folder, so they no longer show up in git status or diffs for folders that are code repositories.

Removed

• Editor Suggestions - the unfinished note autocomplete (ghost-text as you type) and its Settings toggle are gone. The Mira-driven inline ghost (Continue and Rewrite) stays.

Fixed

• Clicking in a note no longer paints a stray selection, double-clicks near inline-code no longer over-select, and clicked carets land exactly where you point, even inside a code chip.
• Line numbers line up precisely with their lines again, and clicking a line number selects that whole line.
• A streaming reply stays visible once it grows past the bottom of the view instead of scrolling out of reach.
• Plan task-detail windows keep the size you drag them to, and agent brief fields grow to fit what you type.
• The Extensions list and detail pane scroll reliably in both side-by-side and stacked layouts.
• The startup splash clears correctly even when the window opens hidden.
• Fewer misleading "retried" labels in the activity view.

Added

• A calmer failure screen - in the unlikely event Undra can't start, you now see a short plain-English message with a button to reinstall from undra.com/download, and a note that your notes and settings are untouched - instead of a technical error box.

Changed

• Releases prove they launch - every release build is now started once automatically, on a clean profile, before it's allowed onto the update feed. A build that can't boot can't ship.

Fixed

• Fresh installs start clean - new installs no longer show a placeholder "My Workspace" flagged Missing; Home opens with New workspace, Link existing folder, and Import front and center. Installs that already had the dead placeholder clean it up automatically on next launch.
• Removing your last workspace now leaves that same clean Home state instead of recreating a placeholder.

Fixed

• App failed to launch after updating to 0.6.50 - a library used by the new workspace-export feature was missing from the packaged build, crashing the app at startup. It ships correctly now, and releases are checked automatically for this class of packaging mistake from here on.
• If you were affected: download the installer again from undra.com/download and run it. Everything you had (notes, workspaces, settings) is exactly as you left it - only the app program itself was broken.

Added

• Capture notes, not just tasks - the global quick-capture popup can now save a note as well as a task: pick the kind per capture, choose the folder a note lands in with a searchable picker, and set your preferred default in Settings.
• Hand a capture to Mira - Ctrl+Enter in quick capture sends what you typed to the assistant instead of saving it, for when a thought is really a request.
• Cursor as a provider - connect a Cursor subscription and chat runs on its Grok 4.5 model with the full agent toolbox: it can read and edit your notes, work in connected folders, and use every Undra tool.
• Rename workspaces from Home - right-click a workspace in the Undra Home list to rename it in place.
• Live preview for house rules - the AGENTS.md instructions editor on a connected folder now renders its markdown live as you type.

Changed

• A leaner quick capture - task details (due date, importance, plan) tuck behind a More toggle, so by default the popup is just a box you type into.
• Steadier multi-window Undra - pop-out windows recover after a crash, follow your theme, dim their frame when unfocused, close with Ctrl+W, and clean up properly when you switch workspaces. Keyboard shortcuts now act on the window you're actually using.
• Mira double-checks names - when you mention a note or item by name, the assistant resolves exactly which one you mean before acting, and asks instead of guessing when it genuinely can't tell.

Fixed

• Every chat tab keeps strictly to its own conversation - fixed a rare case where busy parallel chats could bleed context between tabs, and chat documents now save and reload their transcripts reliably.
• Resizing a dashboard widget now tracks your cursor exactly instead of lagging behind it.
• Codex turns start reliably on Windows and use the model you saved instead of a default.
• Chats that aren't attached to a folder now know about your connected folders, so folder questions work without setup.
• A dock chat's unread dot clears when you open it.
• Files other apps moved into your workspace trash now show up there correctly.
• Undra Home polish: Customize stays above the page content, and removing a never-opened workspace from the list works.
• Packaged builds no longer respond to developer shortcuts (DevTools, raw reload).

Added

• Turn any note into a skill - flag a note as a skill and its content becomes standing instructions for Mira, with a short description that tells it when to reach for them. Skills live in your folders like any other note, get a badge in the sidebar, and are edited in the normal editor - the separate skills system is gone.
• Import skills from Claude Code - if you already keep personal Claude Code skills, an opt-in importer copies the ones you pick into Undra as skill notes, ready to edit.
• Connect outside tools - Undra now speaks MCP, the open standard AI tools use to talk to other apps. Connect a server (say, Blender) in Settings and your chats and agents can use its tools alongside Undra's own - or just ask Mira to set one up for you. Anything that would run a new program shows you the exact command and asks first.
• House rules for connected folders - give any folder portal an AGENTS.md instructions file, written right inside Undra. Whenever the assistant works in that folder, it follows those rules - coding conventions, tone, things to never touch.
• Take a workspace anywhere - export a workspace to a single .undra file, choosing which connected folders come along. Double-click a .undra file on another machine and Undra shows you what's inside and asks before importing.
• The graph sorts itself into clusters - related items now pull together into distinct colored groups with real space between them, so the shape of your thinking is visible at a glance. Hovering spotlights a neighborhood, edges stay calm until you focus, and graphs with thousands of items pan and zoom smoothly.
• A project journal on every portal - connected folders surface the running notes Mira keeps about that project, one click away.
• Pop-out windows survive a restart - editor windows you tear off into their own window come back after you quit and reopen, tabs intact.
• Live preview in the Agents window - when an agent works on a web project from the pop-out window, the running page opens in a preview beside the chat and reloads as files change.

Changed

• One assistant: Mira - the three built-in specialists from recent releases are folded into Mira. Big jobs still fan out to a team of helpers behind the scenes; there's just no roster to manage. Agents you made yourself are untouched.
• Dashboards the assistant can really build - ask for a dashboard and Mira now works from ready-made templates, adds single-number stat tiles, and keeps relative dates live. An empty dashboard offers to generate itself, and adding the first widget no longer fails.
• A quieter start - agent pickers, @-mentions, and the sidebar AI section stay hidden until you actually have agents or start using AI, so a fresh workspace isn't covered in machinery.
• Undra Home polish - workspaces without a cover image get their own generated topographic cover, cards go borderless with cleaner hover states, and each workspace has a right-click menu (open, favorite, export, reveal on disk, remove).
• Writing lists in chat - Shift+Enter continues a bullet, numbered, or checkbox list in the message box instead of just adding a blank line.
• Retired - the workflow item type and the standalone calendar page are gone. Plans keep their built-in calendar view, and scheduled work lives in routines.

Fixed

• The graph no longer flickers while you zoom or pan, and selection stays steady during camera moves.
• Large workspaces (thousands of items) open quickly again instead of hanging at the loading screen.
• Dashboards recover from corrupt data on their own, and widgets on plain dashboards are readable in dark mode again.
• Canvas controls and plan calendars follow your theme and accent color instead of fixed colors.
• Tearing a tab off into its own window (and putting it back) never loses or duplicates the tab.
• Opening a .undra file works even on installs where Windows had lost the association.
• Fewer false "blocked tool" warnings from the assistant, and clearer errors when a folder path is slightly malformed.

Added

• Portals are a place to build, not just read - connect a folder elsewhere on your disk, or make one inside your workspace, and Mira can read, search, and edit its real files. Preview any file in-app (images, text, Markdown, HTML), and run and watch a project's dev server in a preview tab that live-reloads as files change.
• An assistant that can see its work - for web and portal projects, Mira can open the running page, look at it, and fix what's off, instead of editing blind.
• Attach a folder to any chat - a Scope pill points any conversation at one or more portals, so the assistant is grounded in those files without you opening the folder first.
• Search everything at once - one search across your notes, plans, connected-folder code, and past chats.
• Steer a reply while it runs - send a message while Mira is mid-answer and it folds your steer into the current turn instead of cancelling and restarting.
• Live to-do lists - for a multi-step job, Mira shows a checklist that ticks off as it works.
• Long-term memory - the assistant can carry a small set of durable facts between chats, with your approval on what sticks.
• Background jobs - longer tasks run in the background with their own panel, so you can leave them going and come back. Unattended runs are opt-in (off by default).
• Auto effort - the effort dial gains an Auto setting that sizes each message for you.
• Local models can use tools - Ollama and LM Studio models can now call Undra's tools, not just chat.

Changed

• A rebuilt Undra Home launcher - a simpler two-column launcher, the app logo in the activity bar with window controls in the corner, drag-reorderable workspaces with cover images, and Ctrl+K from anywhere. (The old Home chat panel is gone; chat lives in its own panel and window.)
• Clearer portals - "External Portal" (a folder elsewhere) and "Workspace Portal" (one inside your workspace), with connect and disconnect replacing mount and unmount, health states, and re-pointing when a folder moves.
• A clearer assistant - a live table when several agents work at once (each one's task, tools, tokens, and time), plainer previews of what each tool did, and a more accurate context and cost meter.
• Every agent gets the full toolbox - not just Mira, with per-agent opt-in for what each can use.
• Undo survives a restart - the one-click undo for AI changes now works even after you close and reopen the app.
• Draft routines in plain English - describe a scheduled routine in a sentence and Undra drafts it, plus a one-time schedule option.

Fixed

• Stop now keeps the partial answer on every provider, cancels queued turns, and no longer double-reports errors.
• Chats no longer cross transcripts between windows or tabs.
• The graph updates smoothly when you add or link an item instead of redrawing from scratch.
• A missing workspace folder prompts you to relink it instead of failing silently.
• Wikilink typing, the note selection-toolbar popovers, and caret behavior around links.
• Mid-generation stalls are detected and recovered on their own.
• Security hardening: network fetches pinned to a validated address, secrets stripped from tool subprocesses, tighter file-path scoping.

Added

• Building blocks for agent instructions - a panel of ready-made pieces (tools, skills, follow-up actions) you can drop into an agent's instructions, each one documented so you know what it does. The tokens show which parts are fixed and which you fill in.
• A slash menu while you write - type `/` in an agent's instructions to insert a building block, using the same popover as note wikilinks.
• Set permissions by dragging - flip an agent's tool permissions on or off by dragging across the rows (or middle-mouse paint) instead of clicking each one.
• Portrait icons for the specialists - Eco, Snow, and Leonardo now have face-cropped portrait icons in your agents list.
• AI assets go to the trash - deleting an agent, skill, or tool now moves it to the trash instead of removing it outright, they come with a default icon, and built-ins can be reset in one click.

Changed

• Refreshed Settings - Settings now uses one consistent card style across General, Sounds, Extensions, and the rest, with an Extensions hub that reflows to fit the window.
• Light mode fixes - cleared up washed-out, muddy panels in light theme across dashboards, the Home screen, Settings, and the agent editor, with better card depth and softer tab shadows.
• One search everywhere - the Home search bar and the command palette now share the same search, highlighting, and results.

Fixed

• Read-only built-in agents no longer accept edits to their name, description, instructions, skills, or tools.
• The agent slash menu no longer steals focus from the editor or flickers near the screen edge.
• New items dragged in the sidebar now land after the last item in a folder instead of jumping to the top.
• The assistant's listed tools now match what it's actually allowed to use, cutting false "blocked tool" warnings.
• Various stability fixes, including the workspace visualization staying in sync when it renders off the main thread.

Added

• Automatic specialist hand-off - Mira now recognizes when a request is a better fit for a specialist and passes it along for you, no @mention required. You get a card showing who took it and what they did.
• Meet the specialists - three built-in agents now come with Undra and appear in your agents list: Eco pulls scattered notes into a clear write-up with source links, Snow reads your plans and notes as data (exact counts, overdue work, status reports, dashboards), and Leonardo turns ideas and messy notes into canvas diagrams.
• Copy any message - copy your own messages, or the assistant's, straight from the chat, plus smaller polish to how messages behave on hover.
• Keep your question in view - an optional setting pins your latest message near the top while the reply streams in, so you can read along without scrolling. Toggle it right from the message box.

Changed

• A livelier activity timeline - the step-by-step view now animates as the assistant works (each step turns from a working cube into a check as it finishes), and specialist steps name the exact notes and files they touched.

Fixed

• Folders the assistant creates while exploring your workspace now stay scoped to the right place.

Added

• A live activity timeline - watch the assistant work step by step: what it's reading, searching, thinking through, and changing. Each step opens up if you want the detail and tucks away once it's done.
• File edits shown as diff cards - when the assistant changes a file, the reply shows exactly what it added and removed, line by line.
• The assistant shows its thinking - its reasoning streams live as it works and stays on the finished reply, so you can see how it reached an answer.
• An effort dial - choose how hard the assistant works on a request: Quick, Balanced, Deep, or Max. Works with every provider, including local models.
• Work with portal files from any chat - the assistant can read, search, and edit the files in a connected folder from any conversation, not just one opened inside that folder, and make several edits in one pass.
• Answers from the real docs - ask how something in Undra works and the assistant draws from the actual documentation.
• Create dashboards by asking - describe the dashboard you want and the assistant builds it.
• Saved drafts, clearer questions, a live cost footer - half-written messages are kept per chat across switches and restarts; the assistant asks a quick question when a request is ambiguous instead of guessing; and a footer shows how long a reply has been running and what it's costing.

Changed

• A refreshed chat - a new welcome card that reflects your workspace, your messages as timestamped bubbles, and a new assistant avatar.
• Clearer parallel agents - when the assistant splits a job across several agents, the redesigned view is easier to follow, and it now double-checks its own work and that anything it cites actually exists.

Fixed

• No more lost tasks - creating or updating a plan with several tasks at once no longer drops any of them.
• Canvas and dashboard content stays put - adding or removing one item no longer wipes what was already there.
• Portal edits show up right away - after the assistant changes a connected folder, the sidebar refreshes immediately.
• A more accurate context meter - it reflects your real last turn instead of a stuck estimate, and no longer inflates from cached data.
• Steadier tools and clearer labels - hardened the assistant's file and web tools, and activity steps now show note titles and filenames instead of internal ids.

Added

• Click straight to what it finds - ask the assistant to dig something up and the notes and items it turns up come back as cards. Click one to open it.
• See the work behind a reply - each answer shows what the assistant actually did to get there: what it read, searched, changed, or looked up. Open any step for more.
• Jump to mentioned notes - when the assistant brings up one of your notes, it's a link. Click it to go straight there.
• Put several agents on it at once - for a big job, the assistant can split the work across a team of agents working at the same time, then pull it all back together. Changed your mind? One Undo reverses everything they did.
• Deep research - ask a big question and the assistant sends out a team to dig in while you watch them work.
• More it can do for you - find things across all your notes, search inside a connected folder, hop between linked items, open or tidy items away, and act on whatever you've got in front of you.

Changed

• A calmer chat - watch the assistant work as it goes, without the clutter.

Fixed

• Switching notes is much faster - opening a different note no longer slows things down.
• Your tags stay intact when you save.
• Top-level items are named correctly, and the assistant is better at finding and reading your notes.
• Settings no longer crashes when you close and reopen it.
• More stability fixes, plus a quicker quit from the tray.

Added

• Focus layout - clear the side and bottom panels for distraction-free editing, then bring them back in one click.
• Optional title bar - prefer a classic window title bar? Turn one on in Settings. It stays off by default.

Changed

• Refreshed window design - the workspace is now a set of floating, rounded panels on a darker backdrop, with the window controls tucked into the top corner and a slimmer, cleaner frame. Editor and panel tabs connect into their panels as rounded cards.
• Polished pop-out AI window - the detached AI window has a cleaner shell and context panel, plus a small avatar that reflects what the assistant is doing.
• Tidier assistant replies - chat answers no longer narrate every behind-the-scenes step, and the formatting is cleaner.
• Undra always uses your Claude subscription - even when an API key is set in your environment, Undra bills your Claude Code subscription instead of the key.

Fixed

• Dragging tabs between the pop-out and main windows, and dropping onto an empty pane, now works reliably.
• You can drag the window again from the panel tab strips.
• The Panels menu responds to clicks again.
• Long-running chats that used to get stuck now recover on their own.
• Various window, tab, and stability fixes.

Added

• Message branches - start a focused side conversation from any chat message without derailing the main thread. The branch keeps the original message pinned at the top; once it has replies, the message shows a "branch replies" count you can click to reopen it. Your "@agent" mentions work inside branches too.
• Queue up chat messages - keep typing and sending while the assistant is still replying. Your messages line up as chips above the box; reorder them, remove one, or "send now" to jump the line.
• One-click Claude setup - install Claude Code and sign in straight from the Providers screen, with real status instead of just "found".
• Note revision history - browse and restore earlier versions of a note.
• Smarter [[ links - a popover suggests notes as you type a wikilink, and offers to create one if it doesn't exist yet.
• System tray menu - a themed tray icon with quick actions, plus optional "keep running in tray" and "launch on startup".

Changed

• Smaller download, simpler AI - Undra now uses the Claude and Codex command-line tools you already have installed instead of bundling its own, cutting roughly 660 MB from the app. No API keys to manage.
• Search runs on your machine - semantic search builds its index locally, so it no longer needs an OpenAI key.
• Cleaner notes - indentation and the indent-guide lines now line up exactly, like a code editor.
• Calmer background windows - an inactive window gently desaturates its frame so the focused one stands out.

Fixed

• Stop now reliably cancels a reply the instant you click it.
• Deleting a folder no longer discards non-Undra files inside it; they are moved to the trash.
• Background and scheduled AI edits can now be undone, like interactive ones.
• Hardened security, plus various startup, tab, and stability fixes.

Added

• Images in AI chat - paste, drag, or pick images into a chat and the assistant can see them. Attach several at once. (Works with Claude models.)
• Chat tabs - keep several chats open in tabs across the top, each with its own activity spinner and right-click menu.
• Chats list in the pop-out window - the detached AI window now has a sidebar of your chats, grouped by date (Today, Yesterday, This week), each with a live status dot. Collapse or resize it to fit.
• First-run AI setup - a quick setup step the first time you open the AI page, with a "set up again" option whenever you want to redo it.

Changed

• Smoother startup: one loading screen with a real progress bar, and the pop-out AI window now opens with its own loading screen and a proper title in the taskbar.

Fixed

• Running several chats at once no longer crosses their replies; each answer stays in the chat that asked.
• Chat history stays in sync across windows. Deleting a chat in the pop-out window updates the main window too.
• The model picker keeps showing your current models even when the online model list is briefly down, instead of dropping back to an old list.
• Chats now show real errors (an unavailable model, a failed turn) instead of a silent "No response."
• The chat view no longer pulls you back to the bottom while you are scrolled up reading.
• Send Feedback works again.
• Various image-attachment fixes: oversized images are downscaled to fit, drop areas behave, and unusual filenames are handled.

Added

• Extensions - add new item types, canvas widgets, and commands to Undra. Manage them in Settings, where each is marked Verified or Unverified.
• Build and share your own - make your own extension, package it into a single file to share with someone, or add one straight from a file.
• Browse community extensions - find and add extensions and agents other people have made.
• Let AI build one for you - describe what you want and the assistant builds the extension for you.
• New kinds of items - Undra is no longer limited to notes, plans, and canvases. Extensions or the AI can add new item types, each with its own editor and saved as plain files.
• Pop-out AI chat - open an AI conversation in its own window.
• Several AI chats at once - run more than one conversation at the same time, each kept separate.

Changed

• Documentation moved online - the guide now lives at undra.com/docs, with a live demo and updates between releases. The in-app Docs button opens it, and the Milestones walkthrough is now a Home-screen tile.
• Community agents install into your workspace and update in one click.

Fixed

• New canvases always keep everything you put on them.
• AI agents now stick to exactly what you've allowed them to do.
• Lots of smaller stability and security improvements.

Added

• Folder Portals - open any folder on your computer right inside Undra, with an AI assistant that can read your files and make changes once you approve them.
• Bring your own AI - works with Claude, ChatGPT, and local models (Ollama, LM Studio), so you can pick whatever you prefer.
• Scheduled AI routines - set an AI agent to run on its own, daily or weekly.
• Sounds - optional audio feedback as you work, fully customizable in Settings: pick a preset, tweak it, or upload your own.
• Automatic updates - Undra keeps itself up to date in the background; just click "Restart to update" when you're ready, and see what changed right here.
• AI agents & skills - create, duplicate, and delete your own.

Changed

• Large workspaces load noticeably faster.
• The AI panel in the sidebar can now collapse and resize to fit your layout.

Fixed

• Your notes and items keep their content safely when you rename or reorder them.
• General stability and security improvements.