What's new

All voices now sound clearer and more premium

• Faster startup and smoother performance, especially on Windows
• Use Space normally in hands-free mode and customize double-tap controls
• Agent Mode remembers tool confirmation settings after app restarts
• Microphone issues now include clearer recovery guidance
• Fixed a bug that could insert dictated text in the wrong place
• VoiceOS does less repeated work in the background, reducing pauses when reading settings and keeping windows in position
• On Windows, dashboard animations and the Apps page now recover correctly after moving or resizing an unfocused window
• VoiceOS can now restart stopped audio capture and selected-text helpers automatically, so a temporary failure does not keep interrupting later requests
• VoiceOS now keeps your selected microphone in sync more reliably when audio devices change
• Typing spaces during hands-free recording works correctly, stopping no longer accidentally starts a new recording, and Escape cancellation is more reliable on macOS
• Agent Mode now follows your enabled language choices more consistently, avoiding unexpected switches to languages you have not selected
• Dictation now spells common AI model names more accurately
• Custom app background tasks can now show rich result cards when they finish
• Links in custom app widgets now open in your default browser

Agent Mode is now 2.4× faster — quick requests fly, spoken replies start sooner, and it puts less strain on your computer

• When creating a prompt in Agent Mode, drawing now starts faster and feels smoother
• Finder searches are now much faster and surface more relevant filesMac
• Dictation now handles punctuation, quotes, and cursor boundaries correctly across 24 more languages
• Agent Mode's double-tap recording and Escape cancellation are now more reliable
• Finder now ranks strong filename matches first and searches inside matching folders for the requested file typeMac
• Finder results now appear sooner, while stale searches and late previews are ignoredMac
• Images captured in Create Prompt now stay attached to the finished prompt
• Simple Agent Mode requests now take a faster route, while complex and research-heavy work still gets the deeper reasoning it needs
• Agent Mode prepares its tools while opening speech is generated, reducing the quiet wait before a spoken answer
• Approved confirmation cards now reach the action that needs to run more reliably
• Writing styles preserve technical text more carefully, apply punctuation selectively, and stay in sync with the right account
• On macOS, dictation now keeps the insertion point correct after paragraph breaks, emoji, and other rich textMac
• On Windows, the top and side overlays now share one cursor update path, reducing flicker and hover jitter
• Payment-failed and subscription-ending banners now lead straight to the right recovery action, in English and Japanese
• New subscriptions can be completed inside initial setup with supported wallet options, with an automatic browser fallback if the embedded checkout cannot load

• Attach screenshots and reference images directly in conversations with the custom integration builder
• Agent Mode now shows your polished request first, instead of briefly flashing the raw transcript
• Screenshots attached to a created prompt remain visible while you review and refine the draft
• Spoken Agent Mode replies and card summaries now play without dropped or repeated lines
• The recording start sound now plays only after microphone audio is actually flowing, reducing false starts
• VoiceOS now releases a stale pre-warmed microphone stream cleanly when your computer wakes
• Initial setup now offers a skip after 12 seconds and keeps its ambient soundtrack quieter
• Screenshots stay attached across revisions to staged Agent Mode cards, so visual context is not lost while you make changes

• X (Twitter) is back in Agent Mode — post, search, read your timeline, and check DMs by voice
• Jarvis now speaks with a proper British butler delivery, and a new Midnight voice keeps the previous Jarvis sound
• Hovering a screenshot in the notch now shows a preview big enough to actually read
• Quick questions stay in front instead of being handed off to a background task you never asked for
• Fixed agent replies occasionally showing raw internal data, and agent voice going silent on Windows after switching speakers or headphones
• X results now arrive as readable cards — your home timeline, profiles, and DM threads laid out as a conversation — with clear confirmations before anything is posted or sent
• X reads and DMs depend on your X API tier, so the integration is marked beta
• Five agent voices are now available: Sky, Dawn, Dusk, Midnight, and Jarvis. Anyone already using Jarvis keeps it and moves to the new delivery automatically
• Screenshot previews now match each shot's shape, flip above the pointer when there's more room, and no longer get clipped by the bottom of the notch window
• In Create Prompt, the circle you draw now appears only in the captured screenshot, not on the frame that flies into the notch

Talk through what you want and circle anything on screen — VoiceOS turns it into a ready-to-paste prompt with numbered screenshots

• Double-tap your Agent Mode key to keep talking hands-free
• Agent Mode is now 12% faster, with more natural spoken replies that start sooner and light up word by word in the notch
• Polished mode is better at recognizing names and terms visible in the app you're using
• Custom integrations can now use Hooks to react to your voice, add context, reshape requests and results, and ask for confirmation when needed
• Reply notifications now open in the notch with an inline message preview and actions
• Integrations that use Hooks must ask for transcript access, and you can turn that access off at any time from the integration's settings
• The custom integration builder now shows every build step, live terminal output, and readable action logs, with clearer setup previews and automatic crash fixes
• On Windows, circling now captures the hovered window correctly without freezing the trail
• Capturing screenshots and drawing circles no longer slows down the rest of VoiceOS
• Spoken Agent Mode replies stay reliable through longer answers, with smoother voice playback and fewer cutoffs
• First-run setup has been rebuilt into one smoother flow, with a clearer intro, language and profile steps, permission guidance, and hands-on practice
• Fixed duplicate reminders after reconnecting, restored clipboard contents when a paste fails, and improved Windows overlay reliability
• Subscription controls now explain trial cancellation more clearly, and eligible returning subscribers can accept an 80%-off offer in one click
• Live support now handles attachments and update feedback more reliably

Double-tap your dictation key to dictate hands-free — no holding. One more tap ends it

• Claude Code and Codex are now on Windows too — still in beta, tell us what you think
• The custom app builder got a redesign — build and refine your app in a simple chat
• Custom apps now run out of the box, no extra installs needed
• Plus three more small bug fixes and improvements
• Studio now begins with a conversation: describe the integration, review the plan, and keep refining it in the same thread instead of restarting a form
• A new Preview shows what the generated app will do before you install it, including the setup fields and actions it will expose
• Build progress, terminal output, and readable debug logs are visible inside Studio, and common runtime crashes can be diagnosed and fixed automatically
• Custom apps include their own runtime, so shared integrations can start on a new Mac or Windows PC without asking the recipient to install Node.js or Bun
• First-run accessibility setup now includes a guided drag-to-grant flow, with clearer feedback when macOS still needs permissionMac
• Japanese dictation now uses VoiceOS's hosted Whisper path, and Polished mode keeps spoken English product names in Latin characters instead of katakana
• Agent Mode can show verified YouTube preview cards, and map results open named places individually instead of losing the selected destination
• The Windows tray icon now stays sharp on scaled and high-density displays

• Fixed dictation failing with connection errors on some networks and VPNs, and it now reconnects on its own
• Fixed Agent Mode running slow on Windows
• Fixed Agent Mode getting stuck on the loading state after canceling with Escape
• Fixed dictations coming back empty when the microphone glitches
• Fixed the dictation shortcut reverting silently when you changed it during setup
• Added an Appearance section with light, dark, and system themes plus live previews for supported notch materials
• Scheduled reminders now appear as a compact timeline, with clearer app identity, times, and actions in the notch
• Claude Code and Codex gained guided one-click setup plus clearer local or synced session controls, attachments, and task status cards
• The app builder keeps its chat after the window closes and can answer questions about a draft without rebuilding it
• Dispatched tasks automatically open the side notch, while cancelled sessions and long-running results reset more cleanly
• Polished dictation filters false-start markers and letter-by-letter garbles, and switches to a fallback model when polishing goes silent
• Screen drawings are captured at a more readable weight and resolution so Agent Mode can understand the area you pointed to
• Help Center live chat was restored, and Gmail actions gained a more reliable direct-send path with model failover

Build anything: describe the app you want and VoiceOS builds it — test it, save it, share it

• Claude Code and Codex integrations are here — they're in beta, and we'd love your feedback
• Agent Mode is now 30% lighter
• Some dictation improvements
• Apps now have a dedicated dashboard section where you can browse first-party and custom integrations, reorder My Apps, and paste an image directly onto an app icon
• Studio can build from a screenshot, keep multiple builds running at once, and show integrations shared with you
• Generated integrations can be shared by link, with a polished install page, per-tool permission choices, and consistent product pages for built-in apps
• Integrations can report live background-task progress directly into VoiceOS, so active work stays visible without blocking the conversation
• Claude Code and Codex tasks gained durable activity cards, editable queued follow-ups, result and prompt views, effort controls, and visibility into external sessions
• English Polished mode now formats speech for readability while preserving contractions, and Japanese cleanup removes stutters, doubled punctuation, and misheard product names
• Connecting an app from the notch resumes the request that needed it, and disconnected apps now explain why they turned off instead of silently changing state
• Agent Mode added stop controls for active work, kept History open while tasks are inspected, and fixed long answers being clipped before they could scroll
• Map answers now use Google Maps, while coding and integration cards carry the correct app branding throughout the notch

• Fixed clicks not registering on the notch and sidebar on Windows
• Fixed reminder cards you couldn't click on Windows
• VoiceOS now tells you when Windows is blocking microphone access, and opens the right setting for you
• The Windows overlay now recovers its mouse-forwarding hook automatically when the operating system drops it, restoring clicks without a relaunch
• Microphone checks no longer report a false muted state, and real capture failures now surface a clear error instead of silently producing nothing
• Reminder and notification windows keep their action areas clickable after expanding on Windows
• Denied microphone access now links directly to the relevant Windows privacy setting so you can restore recording quickly
• This release also hardened overlay recovery around the notch and sidebar after display, focus, or input-hook changes

New Insights tab. See how fast you speak, the time you save, where you rank, and discover your Voice Type

• Fixed dictation cutting off your first words
• Fixed lag and freezes on Windows when VoiceOS captures your screen
• Cancel your trial or switch plans right from the app
• Fixed notifications that couldn't be dismissed
• Insights turns your history into speaking-speed, words-dictated, time-saved, and percentile metrics once enough sessions are available
• Voice Type gives your dictation style a shareable identity card with a redesigned, privacy-conscious summary
• When you ask about the screen, Agent Mode can request screen access by voice and capture the relevant context for that turn
• Hotkey capture starts faster, and macOS allows more time for slow apps to confirm a completed paste instead of marking it failed too early
• Screen recordings now include the trail you draw for Agent Mode, while Windows screen-context work uses less CPU and causes fewer freezes
• Subscription controls correctly distinguish trials and active plans, and notification handoffs keep their dismiss and action buttons responsive

Outlook mail and calendar, plus Apple Mail on Mac

• Long tasks keep running while you work — you're notified when they're done
• With Polishing set to Polished, Japanese dictation formats lists and numbers properly
• Dictation pastes reliably again
• Plus several other bug fixes
• Long Agent Mode jobs can dispatch themselves into the background, show live activity in the side notch, and return the completed result without blocking your next task
• Queued reminders collect into an expandable notification stack, with smoother handoff between the card, side notch, and active conversation
• Outlook support covers both Microsoft 365 mail and calendar actions, while Apple Mail can search, draft, reply, and send locally on Mac
• Short dictations use VoiceOS's hosted Whisper service for a faster path, with processing-ring timing matched to the selected polishing style
• Japanese Polished mode now formats rather than rewrites: it preserves meaning while cleaning list structure, numbers, email recipients, and stray ASR break marks
• Screen context is captured behind one explicit consent, stays disabled when screen access is off, and can be refreshed for each Agent Mode turn
• Windows dictation paste and the Insert action in reply drafts were repaired across more desktop apps
• Custom URL-based integrations once again support remote MCP endpoints and Authorization headers in their setup form
• Background dispatch now preserves the original transcript and closes several races that could lose, duplicate, or prematurely present a result

Reminders, just by asking

• Faster, more accurate dictation, especially on long takes
• Three polish styles: raw, light, or fully polished
• Hands-free starts on a tap again
• Ask for a reminder, scheduled action, or recurring email check in natural language; VoiceOS freezes the instruction and runs it at the requested time
• Reminder cards support links, update in place when a schedule changes, avoid covering an active notch, and stay aware of the notification already on screen
• The new App Store foundation adds manifest-backed integrations, a browse gallery, capability and permission pages, animated previews, and consistent confirmation cards
• Developers can scaffold an integration from one prompt, preview it in VoiceOS, and hand the generated project to a coding agent for further work
• The dictation pipeline now blends fast and high-accuracy speech recognition, then routes polishing by length and the selected Raw, Light, or Polished style
• An adaptive progress ring reflects the active transcription stage, and long Polished dictations use a model better suited to preserving structure over extended takes
• Files attached in Agent Mode can now be saved to disk for real, while location answers avoid pretending an IP estimate is an exact position
• Idle background work was reduced so VoiceOS uses less CPU when you are not interacting with it and allows the computer to sleep normally
• Agent voice speed is adjustable in Settings, and web answers can include multiple relevant image previews instead of a single thumbnail
• Verified students can access academic pricing, and the cancellation flow more clearly explains eligible retention offers

• New Finder integration. Ask VoiceOS to find, open, and organize files on your MacMac
• Changing your mind mid-request now works smoothly, just say the new instruction
• Improved Agent Mode reliability, with more accurate and consistent answers
• Fixed the mic indicator flickering when starting a recording
• Plus more fixes and improvements
• Finder results use a native-style gallery with thumbnails, drag-out support, Open With, editable names for created files, and clear confirmations before local changesMac
• Folder searches now handle large or truncated listings honestly, resolve familiar folders, and enforce stronger path privacy checksMac
• Agent answers can compose richer native blocks for stocks, charts, lists, files, and tool results while keeping typography aligned with the conversation
• The Home screen now centers tutorials in an edge-to-edge rail, includes English and Japanese subtitles, and remembers completed lessons across devices
• Weekly Agent Mode usage is visible in the dashboard and billing settings, with an upgrade card shown only when a turn is actually blocked by the plan
• Every voice turn refreshes its screen view, preventing answers based on a stale screenshot after the visible app changes
• Empty or stalled model rounds now recover through a fallback path instead of returning a placeholder or blank answer
• Integration connection status refreshes promptly after OAuth, and connect feedback appears as a clear toast instead of shifting the settings layout

• Draw with your cursor while talking to show the agent what you mean
• Pick your own keys for the hands-free dictation shortcut
• Fixed the agent sometimes replying in the wrong language
• Plus 9 more fixes and improvements
• Onboarding is now smoother and easier to follow, with a faster practice step, clearer waiting guidance, and better gating around the first Agent Mode experience
• Hold-to-draw follows the cursor continuously without requiring clicks, making it easier to point at a region while you explain what should change
• Reply drafting gained a dedicated screen-aware card with reliable Insert and Enter actions that collapse cleanly after the text is placed
• Conversation context now survives interruptions and barge-ins, while multiple preview images can stay attached to a visual web answer
• Audio capture detects and recovers streams that die during recording, and the clipboard is restored with all formats—not only text—after Agent Mode uses it
• Dictation defaults to prose unless you clearly ask for a list, with tighter cleanup that preserves the speaker's structure instead of over-formatting it
• Plan cards and the cancellation flow were simplified, with regional currency parity and clearer trial and yearly-offer terms
• Image previews, file confirmations, and integration suggestions were visually tightened so important actions remain visible in the notch

• Ask to find nearby places in the notch with Apple Maps
• Agent Mode no longer replaces text when you're just asking about it
• Cleaner dictation with better lists and fewer cut-off words
• Fixed Agent Mode loading issues and smoother audio ducking
• Plus 6 more bug fixes and improvements
• Nearby-place results preserve the exact map view shown in the notch when opened, with better place hours and smoother result browsing
• Selected text remains visibly attached to the Agent Mode turn until the next request, with improved Japanese bold and italic rendering around that context
• Secure Keyboard Entry is detected on Mac and explained instead of making the shortcut appear broken, while trigger-key choices are limited to reliable modifier keysMac
• Native microphone capture stays warm between recordings for faster starts and uses a lossless stop handshake to protect the end of the transcript
• Media ducking now lowers other audio instead of muting it, remains active through spoken Agent replies, and recovers cleanly on Windows
• Windows gained device location support with clear denied-permission guidance, enabling nearby-place requests beyond the coarse network fallback
• Email greetings capitalize the first body word correctly after a newline, and dictation list conversion now balances structure with wording fidelity
• You can hide the top notch, and long prompts collapse behind an expandable fade so the current answer keeps more room

• Find places near you on a live map in the notch
• Dictation starts instantly, no more clipped first words
• Opening an app now launches the app, not its website
• Apple Calendar improvements
• Hide the side notch when minimized
• Agent Mode moved into a responsive top-notch experience that can reveal your transcript immediately, stream answers into an inline result surface, and remain open until you click away
• The side notch gained a liquid-glass history view with click-to-open conversations, selected-session highlighting, haptics, and an Agent shortcut double-tap
• Answers reveal word by word, remain selectable and copyable, and show clearer source details for web and Google workspace results
• Continuous spoken sessions gained more natural language-aware voices, warmer realtime connections, coordinated start and completion lines, and a true mute control
• Apple Calendar moved to a native EventKit backend for recurring and multi-day events, meeting links, all-day events, and more reliable packaged-app accessMac
• Reminders also moved away from AppleScript to native APIs, improving access and reliability for created tasksMac
• A warm native microphone path makes push-to-talk feel immediate, protects the final audio during stop, and lowers background media while you speak instead of muting it
• Agent Mode can open websites in a named browser or the active Chrome profile, while open-app requests prefer an installed native app before falling back to its website
• Files and captured context can be dragged into the notch with clear attachment staging, and pending confirmation cards return from History after a relaunch
• Near-me requests use live geolocation and a map card, with Windows support and an exact-view handoff when the map opens
• Memory gained lockable core preferences, and confirmation settings were reframed as per-integration permission bypasses so action safety is easier to understand
• Windows focus, paste, and window-activation paths were hardened to avoid freezes, while hover and click geometry around both notches was stabilized

• Agent Mode confirmations are now clearer across Gmail, Slack, and Linear
• You can now paste images directly into Agent Mode chats
• Transcript quality is improved with better filtering of known hallucinated phrases
• Plus 12 more reliability and polish improvements across the app
• Selected text is captured at the right moment and passed into the current Agent Mode request without leaking into later turns
• Stacked previews animate more smoothly, expand based on their real contents, and keep confirmation buttons and overlays reachable in compact layouts
• Gmail, Slack, and Linear handoffs gained more consistent button placement, clearer labels, and better transitions into the native app or browser
• Writing-style instructions were restored and tightened, while transcript splitting became less aggressive so longer thoughts stay together
• Known ASR hallucinations are filtered with language-specific heuristics without removing ordinary phrases that happen to look similar
• Connection-gated tools no longer appear usable before their app is connected, reducing integration errors and misleading Agent responses

• In Agent Mode, you can now open Gmail or Google Calendar to review a draft before accepting it
• We've adjusted transcript accuracy. If anything feels worse than before, please let us know.
• Fixed an issue where Agent Mode integrations could throw an error
• Plus 4 more reliability and performance improvements across the app
• Agent Mode reset timing can be customized with localized timeout choices, and selected text reattaches correctly after a session transition
• Empty speech-recognition results skip polishing entirely, while low-quality audio gets stricter fallback handling instead of producing confident nonsense
• Dictated lists, spelled-out letters, and long paragraphs received dedicated formatting rules, including improved Japanese email structure
• Windows dead keys are handled correctly by the keyboard listener, reducing broken accents and shortcut interference on international layouts
• Screen Recording permission for Agent screenshots now follows a clearer recovery path, and screenshot previews no longer flash at the wrong opacity
• Calendar edits survive the confirmation form, and Slack thread previews render the correct message sequence before an action is approved
• Onboarding troubleshooting gained direct support actions so permission or setup problems can be reported without leaving the flow

• You can now bulk add keywords in Spelling
• Mouse buttons can now be used as voice trigger shortcuts
• Agent replies and action confirmations now feel smoother with less flashing
• You can now contact the founders directly from the dashboard
• Plus 4 more improvements across the app
• The new AI-assisted Spelling import accepts a batch of names or terms, turns them into dictionary entries, and presents the result in a compact review modal
• Restored dictionary entries keep their enabled state consistently, including data imported from earlier versions
• Writing-style cards have stronger contrast and more reliable instructions, reducing cases where dictation ignored the selected style
• Cancelling a dictation no longer leaves the Fn trigger blocked, so the next recording can begin normally
• Agent Mode now pastes the user's intended input instead of accidentally inserting the assistant's response into the active app
• Connected integrations are prioritized in setup and raw internal app links are hidden, producing cleaner confirmations and fewer dead ends
• The founder-contact experience was simplified and fully localized, making direct product feedback easier from the dashboard

• Math in Agent replies now renders properly instead of showing raw LaTeX
• You can now select and copy text directly from History sessions
• Agent replies now include more reliable Slack message and thread links
• Plus 12 more reliability and performance improvements across the app
• Onboarding was rebuilt with a clearer profile and typing-speed flow, animated time-savings guidance, persistent answers, and more natural Japanese copy
• History preserves paragraph and list formatting from the original dictation instead of flattening every saved session into plain text
• The macOS paste helper is prewarmed at launch and runs on a stable event loop, reducing beachballs and delays when text is insertedMac
• VoiceOS blocks overlapping dictations and suppresses the early recording sound while another capture is busy, preventing mixed or lost transcripts
• Voice Agent sessions are saved to History as soon as a reply begins, so a relaunch does not erase the conversation that just completed
• Gmail answers include robust, underlined links to the correct message thread with a clear external-open indicator
• Grammar handling now supports accented initial letters and preserves context-aware capitalization when text begins mid-sentence
• Synthetic copy shortcuts no longer leak a literal 'c' into terminals, and collapsed Agent sessions reset predictably after the configured idle period

• Pasted content is now handled more reliably in Agent Mode
• You can now set icons for custom integrations in Agent Mode
• Agent Mode usage is now shown more accurately in your dashboard stats
• Referral links now work more reliably
• Plus 8 more performance improvements and bug fixes
• Google Calendar confirmations now preserve timezone and event duration correctly, while Docs previews match the content that will actually be written
• Live support accepts an attachment in the very first message and uploads smaller diagnostic bundles with clearer delivery feedback
• Selected Agent text copies as clean plain text, and the selected-text chip has clearer visual contrast inside the conversation
• The iOS app and keyboard gained synchronized spelling keywords, stronger sign-in persistence, and more reliable dictation-state handoff
• Audio that was already muted before dictation remains muted afterward instead of being unintentionally turned back on
• Linear team resolution and confirmation controls were hardened so actions do not stall when team metadata is incomplete
• Custom integration icon previews persist across restarts, and their setup surfaces now keep the user's input in focus while showing clearer approval states

• Added "Show floating bar always" setting
• Added a thinking indicator while the agent works
• Agent chat state is now more reliable, including pinned selected text across follow-up turns
• Fixed an issue where interaction sounds could fail to play
• You can now attach multiple files at once in Agent Mode
• The first What's New system was added so returning users automatically see the release summary once after updating, with access to earlier entries
• Custom integrations gained URL-based MCP transport, an improved explainer, actionable connection logs, and clearer error hints when a server cannot start
• Files can be dragged into Agent Mode, and conversation History now includes Agent sessions with smoother collapse and follow-up behavior
• Pinned selected text survives follow-up turns, while Escape reliably clears the current input and starts a fresh Agent chat when requested
• Notion actions gained a schema-aware confirmation card, and Google Places results can appear as a map with rich location cards before you approve an action
• Agent follow-ups respond faster, maintain their context more reliably, and keep custom confirmation interfaces visible until the next voice transcript is ready
• Interaction sounds are unpacked correctly in production builds, and Windows avoids an infinite restart loop when the Visual C++ runtime is unavailable
• Cross-device spelling keywords synchronize automatically, and shortcut recording now accepts more keys without being blocked by a held modifier