NeuralVault 2.0 is here, as a public Beta

Today anyone can download NeuralVault 2.0. We are happy to say that, and we want to be straight about what it is: a Beta. That means public early access. It runs on Windows, you should expect bugs, and you should keep a backup of your vault before you point it at your notes. We would rather put it in your hands now and hear what broke than wait for a version with every rough edge sanded off.

This is the first time the whole thing has been available to everyone: the application, the agents that work for you, the guard that watches what they do, and the vault they work in. It arrives as one program on your own machine. Before you download, read the Terms and the Privacy Notice. The installer asks you to accept the terms, and the download page says so beside the button.

Why we built it this way

Most AI tools ask you to hand over your work and trust what happens next. NeuralVault starts from the opposite place. Your notes live in a vault on your computer, in ordinary files you can open without us. The assistant, Cortex, works inside that vault, and everything it does that matters is something you can see and decide on.

That one idea shapes the rest of the release. An assistant that can write files and run tools is only useful if you are comfortable letting it, and you are only comfortable if you can see what it is about to do, stop it, and undo it. So we spent the time on the parts that earn that comfort, not only the parts that look impressive in a demo.

Agents that ask first

When Cortex wants to do something that changes things, it stops and shows you a card. The card says what it wants to do and exactly which files it will touch. You read it, and you say yes or no. Each file change you approve is snapshotted first, so one control can put those files back. (That control restores files. It does not undo changes made by shell commands, and it does not rewind a conversation, and we would rather say so than leave you to find out.)

Pick a working folder and chat can create and edit files there, so you can ask for a document, a script or a set of notes and watch it appear in a folder you chose. Paste a screenshot straight into chat. If the model you picked cannot see images, NeuralVault describes the picture for it first, so you do not have to change models to ask about a screenshot.

A guard on every action

Before an agent runs a command, writes a file or sends something out, a separate check looks at what it is about to do. It refuses or asks when the action looks like it would leak a key or password, install unknown or look-alike software, run something downloaded without looking at it, or carry out an instruction that was hidden in a page or a file it read rather than given by you.

We do not claim it is perfect, and we publish how it does. We test it against scenarios held back from its rules and keep track of every one it misses. The scoreboard on the 2.0 page shows the results with their dates, because a single number depends on which test you choose. Expect the guard to ask more often than you would like in a Beta, and tell us when it asks about something that was obviously fine.

A vault that is yours to keep

Your notes are yours, so leaving should be easy and losing them should be hard.

Backups and restore. Restore your notes from a backup, or export your whole vault to a plain folder, from Settings, Vault. Nothing is converted and nothing in your vault changes.
A second copy. NeuralVault keeps backups inside your vault folder, which is on the same disk. So it also reminds you, calmly, to keep a second copy somewhere else.
No lost edits. Close the window with an unsaved note edit and NeuralVault asks you first, instead of throwing your work away.
Nothing left running. Quit the application, or press Stop, and nothing is left running in the background.
A tidier Library. The Library can select several files, move them to the trash, and filter by type.

Your own keys, or ours

NeuralVault works with your own API key from Anthropic or OpenAI, or with a local model through Ollama. Nothing is metered when you use your own key, and you never need anything else.

Personal adds NeuralVault routing: 390+ models through one account, with nothing to connect, and a model picker that lets you choose who serves each turn. It is still in development and it is not on sale during the Beta, so there is nothing to buy today. Free and Personal use the same installer. Your account decides which edition you get, so you can start on Free and nothing needs to be reinstalled later.

More ways to connect

Settings, Connectors has a Discover tab that lists trusted services you can connect, next to your own. You choose what Cortex can reach, and the guard still looks at what it does with it.

Your data, your choice

You are asked about anonymous usage data during setup, and it starts switched off. You can change your answer any time in Settings, under Privacy. When you send a bug report, anything that looks like a key, an email address or your home folder is hidden first. Here is exactly what a bug report sends and hides.

What to expect from a Beta

It will have bugs. If something breaks, report it from inside the application. It takes a minute.
Keep a backup of your vault, and keep copying it. This matters most in a Beta.
Updating. NeuralVault checks for updates and installs them. You can also download the newest installer from the download page and run it over your copy. Your vault is kept. Back up first.
The licence check. If the application cannot reach our service for seven days, it locks. Nothing is deleted, and it opens again when it can check in. See the plan renewal guide.
Windows first. NeuralVault is built and tested on Windows. macOS and Linux are experimental: we test them far less, so they may be more unstable. There is no macOS or Linux installer yet.
Windows may warn you. A new installer from a new publisher can show a SmartScreen prompt while it builds a reputation. The installer is signed by Perpetual Technologies AI Ltd.

What is not here yet

Sync across your own devices, an approvals queue for work that runs while you are away, and a command line you can script are not part of this Beta. Business, our edition for teams, is planned: testing opens in December 2026. The roadmap says what is coming and how far along it is, and nothing on it is a promise of a date unless it says so.

The full list

This post is the why. The changelog is the what: every new feature, improvement, fix and security change in 2.0, in plain words.

Thank you for being early

A first public release is the moment the work stops being ours alone. Every bug you report, every thing that confuses you and every idea you send will shape what comes next. Download the Beta and tell us what you find.

2.0.11001NeuralVaultreleased14 October 2026

Every entry for this release, 268 in all, from the same list the changelog and the release emails are built from. The post above is why; this is what.

Highlights

Updates can install when NeuralVault starts, behind a clear Updating screen. You choose this in Settings, About.
Closing the window with an unsaved note edit asks you first instead of throwing your work away.
Restore your notes from a backup, or export your whole vault to a folder, from Settings, Vault.
Agents ask before risky actions and show you exactly what they will touch.
Pick a working folder and chat can create and edit files there.
Paste a screenshot straight into chat. If the model cannot see images, NeuralVault describes the picture for it first.
Settings, Connectors has a Discover tab that lists trusted services you can connect, next to your own.
Anonymous usage data starts switched off, and bug reports hide anything that looks like a key, an email address or your home folder.
Quitting or pressing Stop leaves nothing running in the background.
The Library can select several files, move them to the trash, and filter by type.

Added

Vault: Keep a second copy of your notes and conversations in another folder, such as another drive, a USB drive or a synced folder. A weekly copy is on by default after set-up, and you can load any backup into a new folder without touching your current vault.
Web search: Chat can now search the web with no key and no account, using public reference sites asked straight from your computer. Your own search key is optional and still wins if you add one.
Web pages: When the assistant reads a web page it now gets the main text in far fewer tokens, and can ask for a section on a topic or the next part of a long page.
GitHub: Sign in with GitHub, or use a token or the GitHub command line, from Settings, Connectors. The assistant can read your repositories, and writes such as comments, issues and pull requests are shown to you first. Pull request and issue chips in chat load their title, state and author.
Chat: A thin strip above the message box shows the working folder's git branch, how far ahead or behind it is, and lets you switch branch after a confirmation. It refuses to switch while files have uncommitted changes.
Chat: The assistant can now stop and ask you a question with two to four choices or a text box, and can keep a visible checklist for a long job.
Chat: Type / to see about forty built-in commands in groups, plus your own saved commands from a commands folder in your vault or project. Nothing is ever sent for you.
Chat: Switch between Plan, Manual and Default permission modes with Shift+Tab or by typing the mode name as a command. After a Plan reply, an Approve plan bar lets you go ahead or ask for changes. The mode takes effect even while a reply is running.
Agent shell: The assistant can start a long-running command in the background, such as a dev server, then read its output, check on it and stop it. It is approved like any other command and is stopped when you delete the chat or quit.
Schedule: Schedule a prompt to run later or on a repeat, once, every few minutes or hours, daily or on named days. Manage it from the chat or the command line, and see next run, last result and cost on the Activity page. Runs ask before changing anything.
Agents: Sub-agents can edit files in a project folder, but only for folders you switch it on for, and every edit still asks you.
Connectors: Connectors whose server asks for a login now have a Connect button that opens the sign-in page in your browser. Connected servers stay running between calls, and the assistant can list and read their resources and prompts.
Connectors: Switch a connector off and on without losing its settings or grants, and edit what it runs from its own page.
Vault as a tool server: Other AI apps can use your vault, read-only by default, to search, list and read your notes. Writing is available only if you turn it on.
Model picker: Simple view shows a short suggested ladder of Efficient, Middle and Top models above the full list. Advanced view adds provider modes and a provider table with price, speed, latency, uptime and region. Your last view is remembered.
Models: In Manual provider mode you can pick a backup provider for the same model, and any problem with the main or backup provider is explained in plain words.
Models: Automatic provider modes skip providers that train on your data or whose policy is unknown, by default. An option can also skip providers that keep your prompts.
Custom endpoints: Route chat to a server you own, such as a home server or rented machine, and add several at once. Private endpoints are treated as your own machine, can be free, and can use plain http on your own network only.
Dashboard and Usage: Rebuilt around a summary of your month, with tokens, money, models, a per-day and 12-month table, and CSV or JSON export. Spend by day is an interactive chart with week to year ranges and hover details. A yearly activity grid shows when you used NeuralVault.
Credits: The credit card now forecasts when your credit is likely to run out compared with your refill date, in words, an icon and a colour.
Notifications: Get a desktop notification when a long chat or Orchestra run finishes while NeuralVault is not in front, and click it to open the chat. Choose how long a run must take before it notifies.
Window: NeuralVault remembers its size, position and maximised state. An optional tray icon, off by default, keeps NeuralVault and its schedules running when you close the window.
Zoom: Ctrl and plus, minus or 0 make the whole window bigger, smaller or back to 100%, and Settings, Appearance shows the current zoom level.
Dictation: Choose which microphone to use from a list beside the mic button. Unplugging it falls back to the system default.
Help: A keyboard shortcuts sheet opens with Ctrl+/ and lists every shortcut the app uses.
Updates: A setting to install updates when NeuralVault starts, a re-check every 4 hours while the app is open, and one calm branded update screen with a progress bar, a stage line and the version going from old to new. Updates apply quietly without the Windows installer wizard.
First run: An optional step to connect GitHub, a short question about how familiar you are with AI that gives newcomers a longer plain-words tour, and a button in the model list that explains every column, tokens and providers.
Chat: Each message shows when it was written, and a Jump to newest button returns you to the latest message after you scroll up.
Context meter: The panel opens on a colour-coded bar and a list of what fills the window, such as messages, memory files, tools, skills and free space, with every figure labelled as an estimate.
Safety guard: A new Settings page, Agents and tools, Safety guard, explains in plain words what the guard does, what always asks and what is refused. It uses fixed rules, with no model and no credits used.
Skills: A skill can limit which tools its turn may use, ask for a specific model you can already use, and take what you type after its name as arguments.
Sidebar: A chat that names one of your workspaces gets a tick and a cross to file it, and three loose chats on one project offer to make a workspace. Nothing is filed unless you tick, and Undo puts it back. Turn it off in Settings, Appearance.
Search: The built-in Web search is listed as a tool on the Skills page, and Settings search now finds sections by everyday words such as backup, dark, memory and model.
Notes: Nested lists, task list check boxes, clickable note links, and Ctrl+S to save in the note editor.
Library: Select several rows at once with Ctrl or Command click, Shift click or Select all, then move them to the trash or drag them together onto a folder. You are told what will go before anything moves.
Library: Filter chips (Notes, Images, Pages, Documents, Other), F2 to rename in place, Delete to send a note to the trash, and a right-click menu on every row.
Library: Browse now opens with your pinned files and your most recent ones above the list.
Library: Select a note to see its tags, what links to it and what it links to, plus a Reveal button to show it in your file manager.
Library: See who last changed a note, whether an AI is writing to it right now, and an Activity list of recent edits with a link back to the chat that made each one. A new Changed by AI filter and an AI changes first sort help you find them.
Vault: Export your vault, or one folder of it, to a plain folder you choose. Hidden files, backups and anything that looks like a credential are left out, and nothing in your vault is changed.
Vault: Restore from backup inside the app. The default copies a backup into a new folder beside your vault; replacing your vault is a separate, deliberate step that shows what will change and keeps a safety copy first.
Chat: Paste a screenshot or copied image straight into the message box with Ctrl+V.
Chat: An image sent to a model that cannot see images is described for it automatically, clearly marked as an automatic description. Images are never sent off your computer for a local model.
Chat: Long chats can refresh themselves into a fresh session at a context level you choose (Settings, Long chats). A handoff note carries the work across and the old chat stays readable.
Chat: A small arrow beside the microphone opens a dictation menu showing whether local transcription is ready and where its model folder is.
Chat: Pull request and issue references in a reply are now clickable links.
Chat: The context meter shows how much of your prompt the provider served from its cache, for the last call and for the whole conversation, and says plainly when a provider does not report it.
Chat: The top bar is rebuilt with the chat title menu, one joined model and working-folder pill, last turn cost, a context ring that turns amber then red as it fills, Tasks and a three-dot Chat settings menu.
Chat: Choose a working folder with one clear button and the assistant can create and edit files in it. Settings, Workspace, Working folder lists your folders, and the assistant now tells you how to attach one if it cannot write files yet.
Sidebar: Group your chats into named workspaces and nested folders, each with its own icon, then move chats with a Move to picker or by dragging. Deleting a workspace or folder never deletes a chat.
Sidebar: Date headings (Today, Yesterday, Previous 7 days, Earlier) split long chat lists, and a Filter chats box finds chats by any words in the title or first message.
Sidebar: Select several chats with Ctrl or Command click or Shift click and archive, move, pin, unpin or delete them together.
Sidebar: Hide a workspace or folder you are not using and restore it later from the Hidden groups line. Nothing is deleted.
Sidebar: A Credit card at the bottom left shows what you have left, a bar, when it refills and a breakdown of recent use.
Connectors: Settings now has Yours and Discover tabs. Yours shows each connector with its status and a Check button; Discover lists well-known services you can add.
Connectors: If a connector loses access you see Access lost with a Reconnect button, and you can try any granted tool by hand from the connector's page.
Connectors: When you ask for something that needs a service you have not connected, the assistant tells you which connector to add and where.
Plugins: Install a plugin from a folder, switch it off and on, or remove it from Settings, with no restart.
Orchestra: Runs can now be split into ordered phases when a later step needs earlier findings, and the run view groups agents under Phase 1, Phase 2 and so on.
Dashboard: A Needs a look panel lists the conversations and board cards that have been quiet the longest, with how long ago each was touched.
What's new: After an update, a short What's new list appears if you opt in. The choice is asked once and can be changed in Settings, About.
Model picker: Redesigned around one search box and a Filters button, with Cheap, Everyday and Hard recommendations, one line per model, a detail panel and a Use button.
Model picker: Each model shows a short tag for the job it suits, such as Everyday tasks or Heavy coding and hard problems.
Currency: Amounts, per-turn costs and prices now show as money in your own currency, with a Currency setting and a one-tap US dollars option. Token usage can show an estimated cost.
Report a bug: A short three-step dialog. The build code is always included, everything else is off until you tick it, and you see exactly what will be sent. Anything that looks like a key, an email address or your home folder is hidden and marked.
Providers: Adding a direct provider fills in the right settings when you type the address of a supported service.
Version: About and bug reports now show the full build code, in plain words alongside the number.

Improved

Chat: The fixed instructions sent with every turn are about 20 percent smaller with no tool removed, so more of the window is left for your conversation.
Web search: Searches wait up to 4 seconds for slower sites and long questions are cut to keywords first, so more searches return useful results. A short plain search no longer asks for approval in Default mode.
Approval cards: Redesigned with Accept, Decline and Something else, a clear header and timer, the exact target, and typed answers. Choice buttons are smaller and tidier, and a blocked call explains how to do it another way.
Guard records: The record of a declined or instructed call now sits directly under the tool call it belongs to.
Credits: Managed requests are charged the real cost rounded to the nearest penny, with fractions of a penny carried forward, so tiny requests are no longer free and slightly more than half a penny is no longer over-charged. The credit card, sidebar and Usage page now count spend the same way.
Credits: The balance updates within seconds of a top-up, and the credit bar now appears for accounts that have only top-up credit.
Usage: Tokens from helper calls such as chat titles, image descriptions and summaries, from stopped replies, and from sub-agents on other models are now counted, and totals say 'at least' when a count was missing.
Models: An alias such as 'Grok latest' now shows the same AI score and label as the model it points to, and two missing measured scores were added.
Model picker: Click a provider row to choose it, with Latency as its own column, a compact privacy icon instead of a red badge, tidier spacing and a score range for each tier.
Model picker: Models whose price is set at request time are no longer offered under NeuralVault Managed, and the warning for models without tool support now says exactly what is true.
Chat: A model call that fails for a passing reason, such as a rate limit or a server error, is tried once more, then moved to the next best model if NeuralVault chose it. A model you picked yourself is never swapped.
Chat: A read-only tool that fails for a passing reason is tried again up to twice, and each try shows on its row. Extra tool calls past the three-per-reply limit now come back saying they did not run, instead of vanishing.
Chat: Leaving the page while a reply is streaming no longer loses the reply so far, and an unsent draft survives restarting the app.
Chat: A note path in a reply becomes a link only when the note really exists in your vault, and notes and Tasks now open in a right-hand bar beside the chat.
Chat: The top bar stays readable with the side panel open, and the Send button is a larger 40px target with clear hover, focus and pressed states.
Tasks panel: A finished task shows when it started, how long it took, a short summary and the tools it used. It shows the true running count, says when older tasks are left out, and shows long names over three lines.
Library: Double-click a folder to open it in place, notes open in a viewer and other files in a preview pane, and the old detail popup is gone. Project and Preview pages have a Back to Library button.
Library: 'Changed by' is now recorded for notes written by Orchestra agents, the command line, the note editor and moves, so it is no longer empty for them.
Vault: A calm reminder explains that backups stay inside the vault folder and that a copy on another drive or synced folder protects you if the disk fails.
Vault: A vault folder that went missing is named on the first-run screen. Choosing a different vault now works even after another build had the old one, and the refusal lifts by itself once the other build has gone.
Vault: A move now says when links in other notes could not be updated, a folder's note count is named for screen readers, and long names show in full on hover.
Backups: A vault that shrinks in steps, such as 100, 50 then 25 files, now triggers the shrink warning, and a Keep my smaller vault button ends the warning when you cleared notes on purpose.
Long chats: The automatic refresh into a new session works on NeuralVault Managed, shows one calm line if it fails, and is a Personal and up feature. Settings and the context meter now say a local Ollama model is needed.
Sidebar: The chat box says Search, the Free card is a slim 'Want more? Get NeuralVault' row, a chat with a damaged file shows as a row that says it could not be opened, and right-click or the Menu key opens the Actions menu.
Connectors: A server that never answers while you test it no longer freezes the rest of the app, and a damaged connectors file is never overwritten without your explicit choice.
Engine health: If NeuralVault's background service stops answering for 30 seconds, a banner appears with a Restart button. When the app is closed or killed, everything it started now exits too.
Performance: An idle NeuralVault wakes far less often, and the licence check and update notice sleep while the window is hidden or minimised.
Licence: An open computer now keeps its place on your plan, so a second computer is no longer let in or refused at random. After an update restart you are let back in automatically, and a network problem no longer reads as 'no access'.
Notes and chat: Block quotes keep paragraphs and lists, code, links and note links work inside bold or italic, links that cannot be opened are no longer drawn as live links, and Arabic, Hebrew and other right-to-left text is laid out from the right.
Export PDF: Links print their address, web images print as text and address, long tables and code blocks no longer leave blank pages, long words wrap, and right-to-left text is laid out correctly. The file is named by date and title.
Code mode: Reading a part of a huge log or data file by line range now works instead of being refused, and a file with one giant line is cut to a safe size instead of flooding the chat. A turn that used instructions written in the project, or read text that reads like orders, now says so under the answer and keeps that note when the chat is reopened.
Rewind: Edits and writes made by the assistant to files in an attached project folder can now be undone, byte for byte, along with your notes.
Git: Git commands now refuse more files that should never be committed, pull stays on the branch the card names, and a GitHub write card says which account it will post as.
Dashboard: Cards sit in even pairs, the Spend by day chart is shorter, and the Guard activity card is readable on Light and only offers Change where the Security page exists.
Activity and Audit logs: Runs show their kind in plain words and a closed-with-the-app run reads 'Stopped when the app closed'. Audit logs show Allowed or Denied outcomes, say when only the newest entries are shown, and have a Refresh button and a working Clear filter button.
Orchestra: A restored run names the task it was for, the result scrolls into view when a run ends, a run with one failed agent says so beside Done, and the Plan chip shows a clear message if the plan file has been cleaned up.
Updates: If a silent update fails, the next start tells you and offers to try again, the downloaded file and disk space are checked first.
Bug reports: One clear sheet with a switch and a Show me for each kind of extra detail, a live exact preview and a short report number, plus a Report a bug button on the sign-in screen. The command line and chat routes hide secrets, emails and paths the same way.
Errors: Many failure messages across the app now read as a plain sentence with a code and Copy error details, instead of raw technical text, including notes in an unsupported encoding, unwritable folders and unreadable history.
Accessibility: A large set of fixes for keyboard and screen reader users across chat, Settings, the Vault, Board, Canvas, Desk, the sidebar and menus. Focus is kept where you expect, results are announced, and more controls have clear names, larger click targets and readable colours.
Settings: Appearance, Tools, Connectors, Skills, Plugins and Vault pages were tidied, with clearer wording, better small-window and zoom layouts, and Free no longer shows controls it cannot use.
Skills: A folder named twice is now one library, a skill another library shadows stays readable, and the empty page explains where skills come from.
Board: Searching says how many cards match, and the detail toggle on a card a search opened now works.
Desk: Counts follow Not a note immediately, the list refills, and each button is named for its note.
Windows: The Start Menu link is put back if it goes missing.
Sub-agents and tools: An assistant that cannot use a tool now tells the model to try another way, and unanswered approval cards, switched-off tools and declined GitHub writes explain themselves instead of reading like refusals.
Mac shortcuts: The zoom row now shows Cmd for all three keys.
Library: Browse is now a real table with sortable Name, Kind, Modified and Size columns. Folders open right in the list, files get previews (including images up to 4 MB), and you can drop items onto any folder or breadcrumb.
Library: A note opens in a read-only viewer in place of the browser and your folders and place are kept when you close it. Folders you have read are remembered for the session.
Settings: Each section is now its own page with a grouped list, a sliding marker, arrow-key navigation and search. Appearance, Privacy choices and Remove my data come first, with the rest following.
Settings: The credits, spending limits, plan, billing and version screens are rebuilt with clearer summaries, bars and a single button that says what it will buy.
Updates: Updates now install inside the app with an Updating NeuralVault screen and no separate installer window.
Updates: About names the release (Release notes for 2.0) rather than the channel you are on.
Look and feel: New NeuralVault logo throughout and a transparent app icon.
Look and feel: The loading screen follows your theme, holds the logo a little longer and no longer flashes the wrong colour at start-up. On Windows the window buttons match the Light theme.
Motion: Menus, dialogs, toasts, the command palette and the model picker now ease in and out instead of snapping. Pages and Settings sections fade in, and everything is off if you ask your system for less motion.
Motion: Board cards rise in and fold away, chat rows in the sidebar fold when deleted or archived, queued messages and approval cards fold when they leave, and the Dashboard bars fill and tiles lift when pressed.
Readability: Secondary and quiet text are darker or lighter as needed so labels, hints, timestamps and placeholders meet contrast guidelines on every theme, and small controls are now at least 24 pixels to hit.
Accessibility: Windows high contrast mode now shows edges on buttons, a clear difference between on and off switches, and outlines on selected rows, pages, tabs and the highlighted command palette row.
Accessibility: Keyboard focus now stays put or returns to a sensible control after saving, cancelling, deleting, renaming, running a check, closing a tour, stopping a reply and many other actions, instead of dropping to the top of the window.
Accessibility: Screen readers now hear search results, save and failure messages, toasts, Vault search progress, Orchestra run outcomes and the running jobs list, and headings no longer skip levels.
Accessibility: Menus take the arrow keys, Home and End, Tab closes them, the Library shelves behave as tabs, the Account menu opens onto its first item and model rows can be picked with the keyboard.
Toasts: They no longer cover Send or the message box, wait while you are reading them or while a dialog is open, close with Escape and give focus back when they leave.
Errors: Failures now read as plain sentences with an error code chip and Copy error details instead of raw system text, across Vault, Library, Skills, Connectors, Settings, Usage, projects and Dashboard cards.
Errors: A timeout, an engine that is starting or restarting, or a connection problem now says what happened and what to do, rather than telling you to check your connection when the problem was local.
Reliability: NeuralVault now restarts its own background service when it stops, shows a banner while it does, and tells you plainly when it cannot. Pages that were waiting load again once it is back.
Dashboard: One failed refresh no longer replaces the list of problems and recent runs; the last good read stays with a small note. Cards balance their heights and run rows line up.
Activity: One failed refresh keeps the rows already shown instead of replacing them with an error, and run rows now line up in neat columns.
Vault: The first search after launch no longer makes the rest of the app wait while the index builds, a note saved during a build is never missed, and the page says when it is building the index.
Vault: A search with no match shortens a pasted paragraph in its heading, folders deeper than six levels are shown, and clicking a folder while searching ends the search and lists that folder.
Vault: Skip to search and Skip to the note buttons let keyboard users cross a large vault quickly, and an empty note now says it is empty.
Backups: Settings now says where backups are kept. A vault emptied by mistake can no longer push every good backup out, and the warning says when the newest good backup is being protected.
Chat: Long chats are sent in steps so providers can reuse their cache, and older tool results are kept longer before being folded away. This makes long chats cheaper, and the model sees 11 to 20 recent messages.
Chat: Line numbers in code blocks keep working after a long chat is compacted, and questions about code are answered with a steadier, more precise setting.
Chat: Claude models used through your own OpenRouter key now benefit from prompt caching.
Chat: The top bar title follows a rename or an automatic retitle straight away, and a very short chat keeps its opening line as its title.
Chat: The first-run tour skips the effort step on models without an effort setting, and the empty chat screen no longer warns about Ollama unless you chose it.
Sidebar: The usage card no longer jumps up and back during a reply, and the update question is now a slim banner across the top of the page instead of a card over the sidebar.
Sidebar: Chat rows show a small icon with a different shape for each state, the rail can be walked with the arrow keys, and a failed rename or delete now says so.
Board: Deleting a card by keyboard keeps focus, and the Add a card field takes focus back after each add so you can type several cards in a row.
Skills and Tools: The tables keep every heading and name whole at the smallest window size, and a skill with the same name in two libraries now shows both, with the unused one marked.
Bring your work across: Bring the newest 50 now takes the 50 most recently written transcripts, and a file that cannot be read is reported as such rather than as empty.
Privacy choices: What is sent now lists everything the switch controls, and Settings, API keys explains that keys are sent only to the provider they belong to.
Notes: A refused save now explains in one plain sentence that the note changed on disk, keeps your text, and shows the rescue copy if one was kept. Cancel then shows the note as it is now.
Report a bug: Closing the dialog asks before discarding what you typed, selecting text across the edge no longer closes it, and Send stays on screen at the smallest window size.
Dialogs: Selecting text and letting go over the dark backdrop no longer closes a dialog, and Escape closes one even if focus has slipped out of it.
Command palette: Holding Ctrl+K no longer flickers it open and shut, and it is filtered to what your edition can open.

Fixed

Chat: A new chat whose first reply was saved while you were on another page now shows its usage and is remembered across a reload.
First run: Continue is no longer blocked by a route you never ticked, a greyed Continue says why, and the chat setup screen gets keyboard focus after Create my vault.
Vault: Text pasted at the top of a note that starts with a dashed header block stays visible after you reopen the note, and saving over a file the app cannot read no longer overwrites it unseen.
Vault: Tabbing through Orchestra run retention no longer saves it, a value outside the allowed range is refused instead of quietly rewritten, and Change vault folder refreshes the backup line and the chat list.
Vault: A save that finishes after you switch notes no longer draws the old note in the reader.
Settings: A Windows blocked folder (Controlled folder access) now names the program to allow and offers a folder Windows does not protect.
Desktop: A folder chooser can now open only after you press a button.
Dictation: A helper that hangs is stopped together with everything it started, and unplugging the microphone mid-dictation ends the session instead of leaving it stuck.
Library: Opening any note now shows that note, and the Library filter chips have proper spacing.
Updates: A leftover note for an older version no longer shows a false 'last update did not finish' notice.
Updates: The update progress bar now follows the download smoothly and never goes backwards.
Sidebar: The changelog card no longer covers other controls or pushes the whole app down, and it stays readable at large zoom.
Usage: Billed calls that end with no saved reply are now counted, stopped replies keep the tokens the provider reports, the per-model table prices output at the output rate, and counted helper calls are no longer counted twice.
Credits: A managed request that fails before it finishes now gets its held credit back, a refunded duplicate subscription no longer wipes your plan allowance, and refunds or disputes take back exactly what they should even when events arrive out of order.
Credits: Refunding a proration or last month's invoice no longer takes this month's plan credit, and a request that finishes after the monthly reset no longer carries the old month's credit into the new one.
Checkout: A double click or a second tab no longer opens a second checkout.
Account deletion: Deleting an account that holds paid credit is allowed after a plain warning that the balance will be lost, and your confirmation is recorded.
Bug reports: A flood of anonymous reports can no longer stop you from filing one when signed in, and an office sharing one address is no longer stuck on a low limit.
Password reset: A first password the server refuses no longer makes the second try say the link expired.
Chat: A message sent while Ollama is not running now says Ollama cannot be reached and what to do, instead of an unexplained error.
Chat: A stale 'Not sent' line now clears when you remove the file, the attach button keeps focus at the file limit, and a screen reader hears when a reply is ready, stopped or copied.
Approvals: A remembered 'don't ask again' no longer answers for a question the guard raised about private content or a secret in a file write, and 'allow edits for this chat' no longer covers a file that runs code later.
Settings, Tools: The policy columns no longer cut off words in small windows, and the Appearance selects fit their card at large zoom.
Audit logs: Actor and Action columns no longer split short words in the middle.
Dashboard: The Guard activity chart now labels each bar for the day it ends on, so the last bar is today.
Free edition: Several pages no longer point to features Free does not have, including the Usage this month row, the Library select-all shortcut, the Plugins empty list and the Tools list.
Notes: A long list in a note no longer splits wrongly, and a pasted picture now says it is not supported yet instead of silently doing nothing.
Updates: A restart no longer shows the full Windows installer,.
Connectors: A failed Test connection no longer says 'could not connect' twice.
Model picker: Spam-clicking a price heading no longer duplicates rows, and Show more now reaches every model.
Home screen: If the usage line or activity grid fails to draw, that part stays blank and the prompt box still opens.
Keyboard: Ctrl+, and Ctrl+Shift+B now work while the cursor is in the message box.
Windows startup: Closing the app, killing it or restarting it for an update no longer leaves background commands or connector servers holding your vault.
Notes: Saving can no longer be lost when two saves, moves or a sync update to the same note race each other, and a note whose name Windows would mishandle is refused with a clear message.
Chat: A chat you named keeps its name instead of being renamed automatically a few messages later.
Chat: A turn that cannot be saved is kept beside the conversation, a conversation that will not open is reported instead of silently missing, and two writers no longer overwrite each other's changes.
Chat: A reply that goes silent no longer leaves the window on working for ever, and an approval card left open no longer stops that protection.
Chat: Confirming an input-method suggestion with Enter no longer sends the message, picks a palette row or submits the first-run and sign-in forms.
Chat: The Search the vault chip no longer wipes a question you had already started typing.
Chat: The slash menu and Dictate button no longer claim there are no skills or no transcription when the check itself failed.
Chat: The message box, Board search and bug report box now show the app's focus ring.
Sidebar: Hidden groups no longer appear in the Move to pickers, and the filter box now finds chats inside hidden groups.
Dashboard: Notes in vault, Folders, Runs finished and Run outcomes now agree with the Vault page and say when they cover only the latest 500 runs.
Activity: Rows no longer jump to the top the moment a long-running run finishes.
Windows: The window is no longer left invisible or blank if the first page never loads, a crashed page offers a reload, and Close window, the X and Quit now work after a crash.
Windows: Closing the app, shutting down Windows or choosing Restart to install now asks first if a note edit is unsaved.
Windows: A false window not on screen error is no longer logged when you minimise early, and the update notice no longer sits on top of Send.
First run: When Windows Security blocks the vault folder you now see which folder, why and what to press. Relative folders, reserved device names and names ending in a dot or space are refused with a clear sentence.
First run: A settings file that cannot be read is explained, never switches usage data on, and is kept aside instead of being replaced. A file saved with a byte order mark no longer wipes your settings.
Library: A big folder with a file-type filter now finds every matching file, not just the first page, and double click or Enter opens the new viewer instead of the old popup.
Library: Rows no longer fail to run or open without showing a code, a refused save is followed by an accurate Cancel, and a screen reader hears what a file is when the Kind column is hidden.
Board: A hand-written section in your board file is no longer deleted the next time the board saves, and it no longer leaks into the heartbeat prompt.
Export: An export that stops part-way now says which file it stopped on and how many were copied, never leaves a half-written note behind, and recovers notes whose names Windows had split.
Images and PDFs: Saving a generated image or exporting a PDF can no longer leave a half-written file, and a damaged image is refused. Code block paths and languages now print in PDF exports.
Files: Chat, connectors, API keys, agent settings and the device id no longer treat a file that cannot be read as empty, and saving survives other programs briefly locking a file.
Notes: Renaming or moving a note keeps links written in a different letter case, and wikilink examples inside code are no longer counted as real links.
Handoff: A second handoff on the same day no longer overwrites the first.
Orchestra: A run no longer dies because one model turned out unusable, returning to the page mid-run warns that a run is still going, finished runs show Finished or Failed, and a timeout no longer says Orchestra cannot run here.
Skills: A skills folder that cannot be read says so with a code and Try again, a very long skill name no longer pushes its card off the page, and an empty search on a machine with no skills no longer says none of 0 match.
Command palette: Right-click and middle-click no longer open a row, and choosing a Go to row no longer leaves keyboard focus on the page body.
Toasts: A second, different toast no longer inherits the first one's leftover time, and a malformed event no longer breaks the others.
Report a bug: The tooltip and About page now name Cmd+Shift+B on a Mac instead of Ctrl+Shift+B.
Billing: If two checkouts were paid by accident, the extra payment is cancelled and refunded automatically, and a refund for an erased account no longer fails for ever.
Usage: The by-model Money column no longer shows a hundredth of the real cost, compact balances such as 8.5k always read in lower case, and the Dashboard Tokens used tile now explains why it can be larger than Usage.
API keys: Storing a key for a service the app cannot check no longer claims it was verified, and the Anthropic row now says no key is set rather than that one was rejected.
Code mode: A tool call written slightly wrong now runs, and a model that promises an action then stops gets one chance to fix it.
Checkpoints: A failed or undone write no longer leaves stray files behind, and one old entry no longer stops a whole Rewind.
Backups: Emptying a small vault is now caught by the backup protection, and the shrink warning counts held backups correctly.
Import: Chats brought across from another assistant no longer vanish when their session name contains a colon.
Accounts: A second NeuralVault account on the same computer now gets its own chat setup screen.
Hooks: A hook with a typo in its event name now tells you it was skipped instead of failing silently.

Security and privacy

Safety guard: The guard now checks every tool call in the default mode against fixed rules and can only make a check stricter. It is far better at catching secrets written into website code, unsafe code patterns, risky installs, commands that run later, and private data leaving after outside text was read.
Safety guard: What a chat has read now carries over to its next turns and survives a restart, so a risky step later in a conversation is judged with the full history.
Safety guard: Many more key and token formats are recognised and refused, sample keys are no longer refused, and the guard asks fewer needless questions about harmless look-alikes. It makes exfiltration harder; it cannot detect a poisoned model, and NeuralVault is not a sandbox.
Models: A model not marked as trusted, such as a local or custom endpoint, now asks before every shell command and outbound call, and is never remembered or run unattended. A notice and a Settings list let you mark a model you know as trusted.
Chat images: Attached images now have location, camera, comments and other metadata removed before they reach the model provider.
Connectors: Connector servers and plugins no longer inherit your computer's secret-named variables, apart from the ones you declare on the connector.
Connector sign-in: Browser login tokens are kept in your computer's secure credential store, sent only to that one server, renewed before they run out and deleted when you sign out.
GitHub: Your GitHub token is kept only in this computer's secure credential store and is never shown to the model or sent to a NeuralVault server. Signing out deletes it.
Web search: Search runs on your computer only. No query goes to a NeuralVault server and no third-party search provider sees it. The approval card names every site that will get the text.
Notes: The assistant's note tools can no longer follow a shortcut or link inside your vault to a file outside it or into hidden app data.
Network: Plain http to another machine is refused for custom providers and gateways, and the Base URL form in Settings now agrees with the app.
Network: Hardened how the app handles outgoing requests and addresses.
Secrets: More provider key formats and key-in-address patterns are now hidden in logs, bug reports and error cards.
Bug reports: Reports from the assistant no longer link to your account and no longer say they go to GitHub. All routes hide secrets, emails, home folders, paths and IP addresses the same way.
Telemetry: Optional, opt-in problem counts and a few timing figures use fixed names only, with no message text and no IDs beyond a random install token. Stored times are rounded to the hour, and nothing is recorded while telemetry is off.
Backups: Pruning backups now takes a file lock so two processes cannot prune at once, and a half-written second copy never looks finished.
Git and files: Automatic git commands stay inside your workspace and vault.
Privacy controls: Provider privacy switches are asked in first run, shown in the model picker and in Settings, Privacy.
Account: Deleting an account keeps your confirmation of the credit loss on record, and the 30 day window still lets you undo it.
Privacy: The first-run usage data switch now starts off, so pressing Continue or Skip never opts you in. Remove my data lists what it will remove first, including recovery copies kept beside your files.
Privacy: Settings, Privacy choices and the privacy notice now describe exactly what a bug report and the optional usage data carry, and bug reports go to a private inbox.
Privacy: Bug reports hide anything that looks like a key, an email address or your home folder and mark it, and Settings, Privacy lists the reports you have sent.
Providers: A custom provider can no longer send your API key over an unencrypted connection to another computer; only this computer and local servers such as Ollama may use one.
Web: Hardened how the app handles outgoing requests and redirects, including requests to addresses on your own network.
Approvals: Writes to the assistant's own instructions, the board and the app's settings files always ask, and a remembered allow edits never covers them.
Approvals: Removed the Bypass permission mode. Outbound calls and moves into Done now always follow the approval rules, and the terminal no longer auto-approves them.
Approvals: The approval card for an outbound call now shows the full address.
Code mode: Hardened against project folders that arrive with hostile settings, and instruction files in an untrusted project make every edit and command ask for approval from the first round.
Agents: Hardened agents, checkpoints and plugins so they cannot read or write outside the places they belong, and a revoked plugin stays revoked after a restart.
Shutdown: Plugin and connector processes are shut down when the app closes, and Stop ends the running command of the conversation you stopped.
Secrets: The secret detector catches more key and token formats and no longer slows down on very long whitespace.
Content safety: Stronger detection of instructions hidden in fetched pages, notes and file names, and a tighter set of protections built into the app window.
Updates: Updated several components to clear known vulnerabilities and tightened how calendar subscriptions and risky file types are handled.