Version history

The installer gets the new live.hexart.io look and says up front that this is an alpha release, with the exact clicks for the Windows SmartScreen warning.

• New installer look. The setup wizard uses the new brand art, sharp at every display scale from 100 to 250 percent.
• Alpha notice. A page right after Welcome explains that the app is in active development: features change between versions and some parts are still rough.
• SmartScreen steps. Alpha builds are not code-signed yet, so Windows may not recognise the publisher. The installer and the FAQ now say what to click: More info, then Run anyway, and Keep if your browser asks. Only do this for a file downloaded from live.hexart.io.

Withdrawn: superseded by 0.0.95a. Only the current version is distributed.

Bloom in the Camera Composer gets a shape and a colour of its own: set how far it spreads sideways and up and down, from a round glow to an anamorphic streak, and give the bloom alone a chromatic aberration. Recordings and streams no longer drop frames under load, and the startup card no longer burns out its highlights on some monitors.

• Bloom width and height. Every Composer layer (Source, Background, Final) has Bloom width and Bloom height next to Bloom size, as multiples of the round bloom. A wide, flat bloom draws a horizontal streak through every light, like an anamorphic lens; a tall one makes vertical flares. At 1 and 1 the bloom looks exactly as before.
• Bloom aberration. Splits the glow into colour, a bluish core with a reddish edge, and pulls its colours apart toward the frame edges like a real lens. Negative values swap red and blue. Only the bloom is affected; the picture itself stays sharp.
• No cost when unused. Both run on the GPU inside the existing bloom pass. Neutral values add no work, and Setup AI and MCP agents can set all three like any other layer control.
• No more dropped frames. Recordings and streams keep a steady frame grid at the output rate, and NVENC encodes in a pipeline instead of waiting on each picture, so a busy frame no longer turns into a skipped or doubled one. Live encoding never holds pictures back for B-frames or lookahead.
• Startup card. The card shown while the app loads, and in About, no longer shows clipped, burnt highlights on some monitors (they never appeared in a screenshot). Its light now covers the desktop behind it instead of adding past white.

Withdrawn: superseded by 0.0.94a. Only the current version is distributed.

VST 3 plugins run live on every audio channel. Setup AI can work through Claude Code, Codex, GLM and other coding agents, using the app's own MCP tools. A PLATE badge shows paid background requests that wait for your click. This release also carries the 0.0.91a fixes, which were not distributed on their own.

• PLUGIN RACK. Load VST 3 effects as inserts on MIC IN (the raw microphone, before noise reduction), MIC, ROOM, MUSIC, SYSTEM, DUB, SFX and MASTER. MASTER keeps a safety limiter behind the chain. Up to 16 plugins per channel. Drag cards to reorder them or move them to another channel (Ctrl copies). Each slot has power, dry/wet and a sidechain from any channel. Each plugin opens its own editor. BYPASS ALL can sit on a hotkey. Open the rack from the FX button under any LEVELS strip.
• Plugins cannot take the show down. The scan runs in helper processes, so a plugin that crashes or hangs while loading lands on the FAILED list. A crash during processing silences that plugin, not the app. The voice chain's latency is subtracted from the A/V offset of recordings and streams. Plugin states save by themselves.
• Setup AI on your coding agent. Pick an ENGINE in the Setup AI window. Presets cover Claude Code, Codex, GLM (Z.ai), Gemini CLI, Qwen Code, Kimi, OpenCode, GitHub Copilot CLI, Cursor and Goose. Any other agent that speaks the Agent Client Protocol works too. The agent works through the app's MCP server, so it sees every setting and action, and risky calls still wait for your Allow. It signs in with its own account or key. GLM takes a Z.ai key kept in the app. With an agent picked, push to talk goes to the agent.
• The agent stays inside the app. App tools run at once. The agent's own shell, file and web tools are refused unless you turn them on, and then each call waits for Allow or Deny. An idle agent closes after 15 minutes and resumes the conversation when the agent supports it. Closing the app ends every agent process.
• PLATE badge. A paid background request from the studio, Setup AI or an MCP agent can wait for your approval. The deck status bar then shows PLATE with the cost, or AI OFF in red when AI is switched off. A click opens Background plates with Generate and Decline on top.
• Fixes from 0.0.91a. Closing the app during a recording no longer hangs. STOP with the transmission delay on always ends the take. COLOR GRADE works again after its gamma or gain was set to 0, and the gamma fields refuse 0. Local servers free their ports the moment they stop.

Withdrawn: superseded by 0.0.93a. Only the current version is distributed.

A hotfix for 0.0.90a. Closing the app during a recording no longer hangs. STOP with the transmission delay on always ends the take. COLOR GRADE works again after its gamma or gain was set to 0. Local servers free their ports the moment they stop.

• Closing during a recording. Quitting the app while it recorded could hang for good. The MCP server waited for the recording encoder to exit, so the app never closed and the take was not finalized. The app now closes in about 3 seconds with both files finalized, and a shutdown guard covers every step of closing.
• STOP with the delay on. With the transmission delay on and the automatic outro off (or its clip missing), STOP was swallowed and the recording went on. STOP now ends the take after the delay, so the file holds everything up to the moment you pressed it.
• COLOR GRADE. An output grade with gamma or gain set to 0 did nothing and could not be fixed from the sliders. Such values now load as neutral curves. The gamma fields refuse 0: 1 is neutral, above 1 is lighter. Gain, gamma and lift each get a one-click reset to neutral. The camera grades get the same protection.
• Local servers release their ports. Programs the app starts, such as the recording encoder, no longer inherit network sockets. This covers the BACKDROP SCREENS server, the browser extension receiver and the Twitch chat connection. Changing the BACKDROP port or LAN access while recording no longer freezes the deck until the recording ends. A closed port is free again at once.

Not distributed on its own: its fixes ship in 0.0.92a. Only the current version is distributed.

An AI agent can now run live.hexart.io. A built-in MCP server gives Claude and other agents every setting and every action. Each change shows up in your deck. Anything risky waits for your Allow. Setup AI builds whole Camera Composer layouts with generated backgrounds. The settings search understands plain commands. General presets save and load the entire app in one file. You can change the transmission delay while live.

• Built-in MCP server. Starts with the app, listens on 127.0.0.1 only and requires a token. An agent can search, read and change any setting (validated, with undo), and run every action. It can screenshot sources and outputs, and read status, logs and performance. Every change appears as a notice in the deck, never on the program output. Going live, ending a recording, paid AI, lowering privacy and saving credentials wait for your Allow or Deny, and credentials are never sent back. Copy connect command in SETTINGS > CONTROL > MCP SERVER copies a ready command for Claude Code.
• Setup AI builds layouts. Describe the scene and Setup AI plans it in the Camera Composer: slots, cameras, backgrounds and a preview. The same plan can switch the director mode, apply a theme, pick the screen sources or change a setting you name in plain words. Backgrounds start with a still from your camera. An image edit keeps the framing and light, with the background darker than your face. An optional seamless video loop adds ambient sound. Every paid generation shows the model and price first and runs only when you click.
• Settings search that understands commands. The search engine finds settings from plain words, typos, half words, synonyms and numbers with units (60 fps, -6 dB, 1080p, 9:16). It also finds actions, hotkeys and Composer pages. Type "make the mic louder" or "turn off subtitles" and it points to the setting with a suggested value. A query takes about 0.04 ms.
• General presets. Save the whole configuration, every setting and the Composer layouts, to one file and load it as a complete profile. Credentials stay out unless you choose otherwise. Device choices apply only when that device exists on this machine. Loading first shows the differences, keeps an automatic backup and tells you what needs a restart. A hotkey steps through your presets.
• Change the delay while live. The transmission delay slider now works during a stream, a recording or NDI. Delay +0.5 s / Delay -0.5 s can sit on a Stream Deck or a hotkey. The change happens under a short Changing delay card with the program sound muted, so viewers never see a jump.
• Delay without desktop lag. The delay's memory budget counts what every app uses on the graphics card and keeps a reserve for the desktop. A long delay no longer makes the whole desktop stutter. Monitors share one frame pool, which cuts the memory a 4K60 delay needs by almost half, so the same card holds a longer delay. If the card still runs short, the delay steps down under the cover card instead of dragging the system.
• Monitor switching under delay. Switching the captured monitor with the delay on no longer stutters. The program shows the old monitor until the delayed picture reaches the switch. Click, key and scroll sounds play exactly when the delayed program shows them.
• Background video sound. Looping background videos and clip playlists in the Camera Composer can play their own sound on a separate Background video bus. It has its own volume and meter, ducked under your voice like music. Only the camera on the program is heard. Off by default.
• AI commentator memory. The commentator remembers facts about you, your viewers, running jokes and past events across streams, picks what fits each line locally and avoids repeating itself. A memory browser lets you search, pin, edit, add or forget entries. In a 10-minute test it used 85k tokens instead of 210k.
• Model pickers. Every AI model field gets a Choose button with live lists from OpenRouter, ElevenLabs, OpenAI and Google. These show price, capabilities, a recommendation and the retirement date. A retired model switches to its successor by itself. ElevenLabs shows only the voice sliders the chosen model supports and remembers them per model.
• Fixes. Cinema mode windows no longer pulse when nothing moves. Cropped or rotated frames no longer stretch their edge pixels when the crop reaches past the source.

Withdrawn: superseded by 0.0.92a. Only the current version is distributed.

Background videos become a playlist that reacts to your audience. The Camera Composer rotates any number of clips, plays a chosen clip when someone follows, subscribes, donates or cheers, and blends clips with a crossfade. Save a whole Composer setup as a preset and recall it with a hotkey. The new LIVE EVENTS section brings in Twitch, YouTube, Buy Me a Coffee, Streamlabs and StreamElements events.

• Background clip playlist. A Video background now takes a list of clips, up to 64 per camera, added several at a time. They play in order, at random by weight (a clip never repeats back to back), or hold on one clip until something triggers the next. Each clip sets how many times it loops per turn, whether it takes part in the rotation, and its own Play button.
• Clips that react to events. Give a clip up to eight triggers from one platform or any, with a minimum amount. Triggers include follow or YouTube subscribe, paid subscription, gifted subs, membership, donation, bits, Super Chat, raid, gift, share or like. Tiers work as expected: a 60 donation plays the clip set for 50 and up, not the one for 5. A triggered clip takes over for its loops. The next ones wait in line. The rotation then resumes the interrupted clip where it left off.
• Transitions. Choose Instant cut, Overlap or Wait for the loop to end, for the whole list or per clip. Overlap is a crossfade of your chosen length, timed to end exactly as the outgoing clip finishes its loop. The next clip opens ahead. The current one stays on screen until the new one has its first frame. There is never a black frame between clips.
• Hotkeys for clips. Any clip can have its own global hotkey, recorded the same way as in the CONTROL tab. It works with the app minimised; a key combination already taken is marked.
• Composer presets. Save a camera's entire Composer setup as a named preset: cutout, layers, filters, placement, masks and the clip playlist with its triggers. Pick it from a list, rename or overwrite it. Give it a hotkey that applies it to the camera of your choice.
• LIVE EVENTS. A new section in OUTPUTS collects audience events. Twitch bits, subs, gifted subs and raids, and YouTube Super Chats, Super Stickers and memberships come from the LIVE CHAT links. Buy Me a Coffee support and memberships arrive with an API token. Streamlabs and StreamElements add follows, donations and alerts from the platforms linked to them. Every source shows its status, tokens stay masked and out of logs, and Send test event checks your triggers without a live stream.

Withdrawn: superseded by 0.0.90a. Only the current version is distributed.

The Camera Composer now places its layers with Corner Pin, position and scale for the camera and background. Light is no longer clipped between layers. There is a Tone section and lens blur with real bokeh. The chroma key gains shadow and highlight keying, noise reduction, temporal smoothing and matte cleanup. Shaders compile in the background, every Composer setting is searchable, and sliders keep values typed past their range.

• Corner Pin and placement. Source and Background each get scale, position and a four-corner pin, even far outside the frame. Place camera and Place background work on the preview. Drag a corner to pin it, drag inside to move, turn the wheel to scale around the pointer. Double-click a corner to reset it. The background shows around a shrunken camera even without a cutout.
• Light without clipping, Tone. Layers add up in floating point and are clipped only once, at the very end. A bright background or bloom no longer flattens to white early. Every layer has a Tone section: exposure, black and white point, shadows, highlights and highlight roll-off. Auto exposure can meter the Final layer, entirely on the GPU.
• Lens blur with bokeh. A disc-shaped defocus for any layer. On the camera it blurs the finished cutout, masks and edge cleanup included; bokeh highlights bloom only from blown-out lights.
• Chroma key refinements. Screen shadows and Screen highlights key the darker and brighter parts of an uneven screen while neutral blacks and whites stay. Noise reduction steadies the key without softening the picture, and Temporal smoothing calms it at any frame rate. Matte cleanup adds Clip black and Clip white, Shrink / grow with the same reach in every direction, and Edge blur. Spill offers three methods (Average, Gentle, Strong), a reach from the edges to the whole person, and Luma restore.
• No shader stalls. Shaders compile on background threads. Switching the Composer to the person refiner on a fresh start used to hold the picture for up to 6.7 s. The first frame now takes under 4 ms. A broken shader edit keeps the working shader live.
• Composer settings in search and Setup AI. All 152 Camera Composer settings appear in the settings search with their tab and section. Open jumps to the right tab and highlights the row. Setup AI can read and change them too.
• Sliders keep typed values. Ctrl+click a slider and type a value past its range: it is kept exactly as typed and used. Such sliders are drawn violet, and their tooltip shows the slider's range. Typed values are stored exactly (1.18 stays 1.18) instead of being rounded to the slider's display. Hard limits remain only where going further would break the picture or stream. These cover opacity 0 to 1, zoom of at least 1, encoder limits and the minimum interval between paid AI calls.
• Fixes. The director's gaze side threshold is now saved with the other director settings. The settings search shows the real range of about 150 more sliders, and Setup AI keeps its values inside it. Auto exposure uses its faster speed when darkening too, as its tooltip says.

Withdrawn: superseded by 0.0.89a. Only the current version is distributed.

A startup crash fixed, and the first public release of everything from 0.0.86a. The Composer's person cutout is rebuilt from scratch: it learns the empty room behind you, so dark hair on a dark wall holds. Curved masks and per-layer filters are new, every Composer control has its own reset, and an eyedropper picks colours from the camera.

• Startup crash fixed. With YouTube Music or a browser layer on, the app could close within a second of starting. This affected launches from Start menu search, the Run box, a pinned program or an administrator terminal. Windows reported an old Windows version to the app there, and the embedded browser stopped on it. The app now declares the Windows versions it supports. A fallback keeps the browser running even when a compatibility mode is set by hand, and the self-check guards against the problem coming back.
• Person cutout, rebuilt. The Composer learns the empty room behind you and compares every frame with it at several scales. A structure test ignores lighting, so monitors and LED strips no longer count as the person, and brightness is normalised block by block. Dark hair against a dark wall, a black hoodie in front of a wardrobe and fast motion stay put. Edges are finished at the camera's full resolution. It costs under a millisecond of GPU time per 1080p frame.
• Matte quality. Quality (about 960x540 working grid) for the finest hair and dark-on-dark edges, or Performance for about half the GPU time.
• Edge integration. Subject isolation keeps the main person and drops stray patches of the room. Edge colour cleanup removes the old room's colour from hair and soft edges. Light wrap and rim glow place the cutout inside the new background.
• Curved masks. Keep and remove masks take Bezier curves. Click an outline to add a vertex, double-click a vertex for a curve, Alt+drag to pull out handles. Drag inside to move, and undo with Ctrl+Z.
• Layer filters. Source, Background and Final each get a grade (gain, gamma, lift, contrast, brightness, saturation, vibrance, sharpen), plus bloom, glow and soften. The Source layer adds auto exposure metered on the person, entirely on the GPU. A neutral layer costs nothing.
• Reset on every control. Each Composer parameter has an R button that restores its default. It is greyed out while the value is untouched, so you can see at a glance what you changed.
• Colour eyedropper. Pick next to Key colour and Background colour switches the preview to the camera picture. Click to take the colour, averaged over a small area so camera noise does not skew the key; right-click or Esc cancels.
• Splash. The start-up splash lights up the desktop additively instead of darkening a screenshot of it, in SDR and HDR.

Withdrawn: superseded by 0.0.88a. Only the current version is distributed.

Rooms grows up: per-guest channels, labels and face framing, a speaker-driven edit, the presenter's screen in the main shot and its own audio mastering. A per-camera Source Composer adds key, person matte, image, blur or looping video behind you. Also included: dual native NDI, ten new Backdrops, a smoother Windows pointer and lighter previews and projectors.

• Rooms: speaker-driven edit. Talk cuts full-frame to whoever speaks and splits the frame when voices overlap. Work, Hybrid and Focus keep their PiP style and animations; extra guests get more PiPs. AUTO gives way the moment you use mouse or keyboard.
• Rooms: every guest on their own channel. Each person has a meter, fader, mute and an editable name label in Audio. Optional Face Tracking per guest (with the guest's consent), analysed on the GPU only while visible.
• Rooms: presenter's screen. The last shared screen takes the main shot with a label; native guests send their pointer so the frame can follow it. Cameras and screens are separate tiles, and guests are told apart by session, not by name.
• Rooms: audio and connection. The mix you send has its own loudness normaliser (adjustable LUFS) and a -3 dBFS limiter. Guest voices are levelled separately and keep your manual trims. Guests pick camera, mic and speakers before they join (or join listen-only). Connect to last room restores the last session and host rights; rooms survive a server restart and the app checks the media bridge for real.
• Source Composer per camera. Replace the background with a colour, an image, a blur or a looping silent video. Use a colour key or the on-device person matte, before delay, grading and the edit. A Composer studio window shows original and result side by side. Draw up to eight keep / remove polygon masks for a desk or keyboard. Better matte edges: finer input scaling, adaptive temporal stability and colour-aware edge fitting on the GPU, with edge, detail and stability controls.
• Camera glow and cutout edges. Glow strength, cutout halo, light wrap and edge softness in one section, with a one-click "no glow".
• Dual native NDI. NDI keeps the native 9:16 resolution. It can send the 16:9 and 9:16 canvases at once as two named channels with the same audio mix.
• Browser layers in 9:16. The same pages get their own position and size on the vertical canvas; drag and scale them in either preview, even across GPUs.
• Backdrops. Ten new animated GPU backgrounds: Prism Atelier, Luminous Weave, Orbital Glass, Contour Dunes, Honeycomb Bloom, Porcelain Ripples, Spectral Ribbons, Starlight Vault, Caustic Lagoon, Kinetic Mosaic. Search, filters and gallery pages are included. Choose which music drives them: the app's music or the full mix.
• Smooth Windows pointer. Monitors use Desktop Duplication instead of Windows Graphics Capture. The system pointer keeps its hardware plane and no longer micro-stutters while the app runs. The window gallery no longer opens capture sessions.
• Lighter previews and projectors. A Performance mode for the PROGRAM and VERTICAL previews (a third of the master rate, never below 15 fps) that leaves recording and streaming untouched. Projectors present outside the panel, pause while minimised or covered, and resize once after a drag. They pace themselves so several on one card no longer starve each other.
• Page Guard for web messengers. Smart Region 1.9.1 recognises WhatsApp Web and Telegram Web, keeps per-tab reports across windows, and the app flags an outdated extension.
• Watermark. The command-line render and the editor now apply the same watermark rules for 16:9 and 9:16 exports.

Withdrawn: superseded by 0.0.87a before it was distributed. All of its changes ship in 0.0.87a. Only the current version is distributed.

INTRO / OUTRO on the touch bar, and a LOOP clip. After a manual outro, the end-card gives way to a clip on repeat. The intro waits for the pass to finish before revealing the show.

• INTRO / OUTRO under your finger. The toolbar's graphic-package button is on the touch bar too (with a clip set). OUTRO goes over the show; INTRO lights while the package is up. The same arrow-into-the-bar / arrow-out-of-the-bar icons are used.
• LOOP. A fourth alpha-video clip next to INTRO / OUTRO / BUMPER (16:9 plus an optional 9:16). After an outro fired by hand the frozen end-card is replaced by this clip on repeat, with its own sound, the programme silent underneath. Build it to end on its own first frame; make it opaque to keep the show hidden.
• The intro waits for the pass. Pressing INTRO over the loop queues it: the pass in progress plays to its end, then the intro reveals the show. The button shows a ring while it waits; press it again to cut at once. START (REC) lifts the loop the same way. The outro on REC stop is unchanged.
• Sets. The INTRO / OUTRO / BUMPER presets carry the loop files as well.

Withdrawn: superseded by 0.0.86a. Only the current version is distributed.

BUMPERS adds an alpha-video clip over the show that hides a cut to a director mode, fired from the touch bar with two taps. Named INTRO / OUTRO / BUMPER file sets are included. A music player card on the touch screen. The operator's monitor keeps the system sound during a pause and under the intro / outro while the audience never hears it.

• MUSIC PLAYER card. The touch-screen projector shows a now-playing card above the bar, bottom-right, with cover art, title and artist. Tap its progress bar to jump to any tenth of the track. It also has PREV / PLAY-PAUSE / NEXT. Works for the built-in music folder and for YouTube Music alike (same transport, same card).
• The level under your finger. For the built-in bed, the card carries ten tappable music-level steps plus −3 / +3 dB buttons, with the value shown. A tap moves the same level as the AUDIO tab's slider, and is saved like the slider.
• Nothing per frame. The card is redrawn only when the track, play state, every 2 % of progress or a level step changes. The cover comes along only then. The badges and the cost panel step up above the bar while it is up.
• Off switch. MUSIC PLAYER card on the touch screen, next to the touch bar switch in the projector options (on by default).
• BUMPER. A third alpha-video clip joins intro and outro: MOV / WebM, 16:9 plus an optional 9:16. It plays OVER the running show with its own sound. Nothing ducks and the take is never touched. At your chosen moment, the director cuts to a new mode under the graphic. A slider ranges over the clip's own length (read from the file when you pick it) and sets the cut point.
• Two taps on the touch screen. BUMPER on the bar arms it: the mode buttons light up with a ring, and the mode you tap next is where the bumper lands. A hotkey / deck action and a PLAY BUMPER button cut to a default target mode instead. An intro or outro on screen wins; a REC stop or a manual INTRO / OUTRO drops a running bumper at once.
• INTRO / OUTRO / BUMPER sets. One PRESET combo in OUTPUTS swaps all six files and the bumper's cut point at once. SAVE AS keeps the current set under a name. DEL forgets it (the files stay).
• The monitor keeps the room. During a pause, the SYSTEM bus (your call, the browser) leaves the programme. The recording, streams, NDI, PROGRAM SEND and room send get none of it. The music bed and co-host stay. Your headphones still hear SYSTEM, raw. The same under an intro / outro and on the frozen end-card: the audience gets the graphic's sound, you never lose the conversation.

Withdrawn: superseded by 0.0.85a. Only the current version is distributed.

The touch-screen projector gets one bar along the bottom edge. Director modes sit on the left, REC / STREAM / MIC / CAM / PAUSE on the right. It is on by default wherever the screen can take it.

• TOUCH BAR, one row, full width. The two-row dock of 0.0.82a becomes a single bar flush with the projected picture's bottom, spanning the screen. Modes come first: AUTO / TALK / HYBRID / WORK / FOCUS / AMBIENT / CINEMA / CONVERSATION / GAME. A divider separates them from the transport. Buttons stay centred and shrink to fit a portrait screen.
• A permanent REC light. REC is always on the bar and glows red while a take runs. One tap starts or stops the recording, no menu, no mouse.
• On by default. The bar was opt-in and easy to miss; it is on now, with the same preconditions. These are a touch digitizer on that screen, click-through and hide-from-capture both on. A one-time migration switches it on for existing configs. The projector options show, per projected monitor, which precondition is missing, and the opacity slider for the bar sits right next to the switch.

Withdrawn: superseded by 0.0.84a. No longer distributed.

Idle pause adds presence layers (cursor, mic, face) and a 0–30 s countdown routed to streams, recording or only yourself. There is a separate resume count for REC, FLASHCUT, and an AI co-host that keeps the room alive while you are away. Indicators are now icons and controls. A touch-screen projector doubles as a control surface.

• IDLE PAUSE block (DIRECTOR tab). Pick what counts as presence: cursor / keyboard, microphone, a face in the camera (with its own “face gone after” delay). You are idle only when every ticked layer says so: “no camera sees me and the pointer does not move” is one tick away. The countdown before the pause is 0–30 s (0 = cut at once). The countdown before REC resumes is a separate knob. DISTURBABLE decides whether the tiniest nudge cancels a countdown or only activity that persists. Streams are never paused; they show the BRB card and the count.
• Where the countdown shows. Streams / NDI, the recording, and/or operator-only (preview + a projector overlay, never encoded). When recording and streaming together, if only one wants the count, the compositor splits the frame. One copy carries the HUD for that group of sinks. This costs one copy per frame, only while a countdown runs.
• FLASHCUT. A white flash lands in the take right before the pause and again on resume. An automatic pause reads as a deliberate flash-cut instead of a freeze.
• The co-host hosts the pause. Stream only: while the take is paused, the AI commentator switches to engagement mode with activation questions, hooks and chat reactions. It uses its own faster cadence, with its ticker forced on. If off, it switches on for the pause and back off afterwards.
• Indicators are icons. Every badge is a vector glyph: a mic, camera, clapperboard for the mode or gauge for the frame rate. The same table draws them on the preview and projector. Text labels are off by default (one-time migration; tick them back in INDICATORS). Badges are controls too: click REC, LIVE, PAUSED, MIC, CAM, SUB, KEYS or the mode badge on the preview to act.
• TOUCH CONTROLS on the projector. On a touch screen with click-through and hide-from-capture both on, a subtle dock is composed over the projected picture. It includes director modes plus REC / STREAM / MIC / CAM / PAUSE. It answers to a finger only while the mouse pointer is off that screen. When the pointer arrives, the dock stands down and the projector becomes click-through again. Hidden from capture by precondition, so none of it can reach the recording or the stream.
• Performance. The pause logic is a pure state machine with its own self-check: layers, disturbable, face dwell, flashcut order, manual resume, feature off. Pause settings are re-read only when the config changes. The touch dock and operator HUD are rasterised only when their state changes. The digitizer probe runs every 5 s, never per frame.

Withdrawn: superseded by 0.0.83a. No longer distributed.

Page guard lifts the frost late, never early. Closing or switching away from a credential tab no longer shows the page for a few frames. Legacy per-field masks are gone: a page is guarded whole or not at all.

• No more exposed frames when you close a guarded tab. The extension reports “page hidden” the instant the browser decides so. The window still shows that page until the next tab or the desktop behind a closed window is painted, one to three frames later. The app used to lift the frost on receipt. Every clear is now fail-closed: the window stays frosted for half a second afterwards, for both live and delayed programme. The hold follows each frame’s capture time, not the wall clock. A page that comes back inside that window is frosted again at once.
• Legacy field masks removed. Smart Region's per-field rectangles before 1.8.0 (a password box, a key cell) are no longer honoured. A missed or mis-measured field was a leak by construction. The page guard’s whole-window mask is now the only browser mask. The “legacy field masks” look section and its seven settings are gone. Every mask draws with the PRIVATE WINDOW look. The deck asks for extension 1.8.0 or newer (1.9.0 ships in this build; no reload needed).
• Lighter when off. With Page guard off, the render thread no longer takes the mask lock. The mask look is resolved once per pass instead of once per mask.

Withdrawn: superseded by 0.0.82a. No longer distributed.

CINEMATIC stops fighting over the framed area. Screen ticks show which monitor your pointer is on. The program send to Meet / Zoom / Discord proves what it sends. Mix-minus lets the far end hear your whole mix, not just the mic.

• CINEMATIC no longer flickers between the player and the window. Auto-engaged cinema mode dropped out on the first cold frame: a buffering dip, ad swap, seek or pointer grazing the player’s edge. It re-armed a third of a second later. The log showed CINEMATIC, then HYBRID, then CINEMATIC pairs 17 ms apart. On screen, the framed area jumped between two crops. Both edges are debounced. It leaves only after 0.8 s of a cold signal and holds at least 1.5 s once entered. After leaving, it needs a longer hover to come back. A manual CINEMA hold is untouched.
• Which screen is that tick? In TRANSPORT, the monitor tick for the screen under your pointer gets a soft green frame and a tooltip saying so. You can tick the right panel for the fullscreen programme without trial and error. Advanced > SOURCES highlights the screen currently on air the same way and marks the pointer’s screen.
• PROGRAM SEND replaces “Virtual microphone”. The section shows the device’s real format: Hz, channels, share mode. A live meter shows what left for the cable by bus: mic, music, system, sfx, guests, voice-over. It also shows an underrun counter and warns when the cable is not at 48 kHz. This proves the far end gets the mix instead of guessing from the master bar.
• Mix-minus for calls: the far end hears everything, not only your mic. When the call app plays participants on the device captured by SYSTEM, their voices rode the send back to them. The call’s echo canceller stripped your music and desktop audio along with that echo. The send has a bus matrix. Untick SYSTEM or any bus for a mix-minus built on the mix thread through its own limiter. With all ticked, it is the finished programme bit for bit, at zero extra cost.

Withdrawn: superseded by 0.0.81a. No longer distributed.

Browser video stops stuttering while the deck runs. Facebook chat shows only comments under your video. Smart Region tells you when it needs a reload. Closing the deck right after STOP no longer looks like a hang.

• Chrome no longer stutters with the deck open. Smart Region's content script walked the whole page after every DOM change, and label lookup grew with page width. A 12 000-cell grid froze the tab for 30 s. It scans once at load and re-checks only changes in idle-time slices of a few milliseconds. It measures field rectangles inside the frame instead of forcing layout from a timer. UI Automation never touches Chromium windows (Chrome, Edge, Electron) on the app side. The first query enabled browser accessibility mode, which makes video playback drop frames for good.
• Facebook chat: the comments under the video, never your private messages. The hidden browser reading Facebook comments is signed in as you, so the page also renders your Messenger dock. The collector lifted every article on the page. A comment is recognised by its permalink to this video and its place in the page, in any UI language. Messenger, notifications, the post body and related videos are skipped. Source status shows how many comments were lifted and what was ignored.
• Field masking tells you when the extension is stale. Chrome keeps an unpacked extension’s old service worker until you click Reload, so updated files silently stopped reporting sensitive fields. Every report carries the extension version. PAGE FIELDS and the extension popup say outdated: reload, with a one-click button. Rejected reports name their reason in the log. A sentence merely mentioning “password” no longer turns neighbours into phantom fields.
• Closing right after STOP waits for your file. The last take keeps finalizing in the background (encoder drain, container close, MP4 remux). Closing the deck in that window shows what is still running and closes by itself once the file is delivered. A Windows shutdown is held with the same reason instead of cutting the recording short.

Withdrawn: superseded by 0.0.80a. No longer distributed.

Privacy masking now follows the actual source texture, while the director, cursor and capture paths settle faster under real desktop work.

• Privacy Guard fixed: protected regions are fully opaque through their corners. They retain their mask through crop, zoom and mip generation, and preserve HDR radiance instead of falling back to SDR.
• Capture stays responsive: screen and window capture producers are paced. Inactive gallery sessions avoid needless GPU work. The native Windows pointer is protected from camera-hand tracking by default.
• Director tracking stabilized: typing and nearby pointer activity trigger PiP avoidance sooner. Head lean and framing filters use time-based smoothing to prevent sudden camera drift and vibration.

Withdrawn: superseded by 0.0.79a. No longer distributed.

A studio-ready factory profile ships out of the box. Make-up gain keeps ducking from costing loudness. The picture shows one frame at a time instead of frames fighting each other. Settings at zero stay predictable.

• A finished studio on first run. The app ships a factory profile with a working studio's look, director tuning, audio mix, hotkeys and overlay settings. These replace bare code defaults. It is a curated asset, not somebody’s config file. Every credential, device binding and personal path is stripped out. The build's test fails if it finds anything shaped like a key, stream URL or absolute path. Reset chips in the panel restore these values too, so “back to default” means back to the studio setup, not back to grey.
• Ducking gives level back, not just takes it. Ducking only ever attenuated, so a show that ducks a lot went out under level. MASTER gained a make-up section. Duck make-up lifts the programme with the bed's actual ducking, so your voice reaches emission level instead of dropping with music. Master make-up is a fixed offset on the sum after the leveler.
• The auto-leveler is finally adjustable. Its lift was capped at a hidden +8 dB, which meant quiet material could never reach the loudness target however long it played. Maximum boost, maximum cut and the two speeds are now controls. The leveler has an on/off switch and starts working on much quieter programme.
• One frame at a time. When a chat card or window card lies over the screen panel, the panel steps back into a backdrop plate. Its ring, HUD, glow and sheen fade to an adjustable amount; the plate frosts slightly. Two fully framed rectangles no longer fight for attention. Chat cards can share the panel’s own perspective, so they lie flat over a flat desk and turn with a turned panel.
• The camera inset stops sitting on the frame line. An inset landing on the screen panel stays inside it. Its margin comes from the theme’s frame width, ring offset and glow gap. An inset beside the panel, or one deliberately riding past the canvas edge, is left where you put it.
• Fixed: theme values at zero could break the picture. A HUD stroke width of 0 made the panel’s inner shadow divide by zero, corrupting the whole panel. Chroma key similarity at 0 broke spill suppression. A lost face poisoned the gaze layer. A cursor scale of 0 blew the cursor up to fill its tile. Zero now means “off” everywhere.
• Camera changes ride the cut. A second camera no longer swaps in mid-shot. It waits for the next cut and arrives with it, with its own minimum spacing and maximum wait. Lean-in zoom reads only the main camera, so cutting to a side camera cannot make it jump.
• The stage grid lives in the camera picture. The grid is composited onto each camera view after colour grading, instead of lying flat over the finished frame. It forms a floor, back wall and optional ceiling in perspective. It includes fog, parallax against your head and a dolly-in as you lean toward the lens.

Withdrawn: superseded by 0.0.78a. No longer distributed.

HDR capture and HDR10 recording arrive with auto-exposure rebuilt around a GPU histogram. Settings move to and from the cloud by hand. The deck's blank window on machines with virtual-display drivers is fixed.

• HDR, per monitor, mixed with SDR. The app now checks the colour mode of each display on its own. An HDR screen is captured in scRGB half-float with its real white level and peak brightness read from Windows. Conversion happens at capture: BT.2390 tone curve with hue preservation, highlight desaturation and gamut compression. HDR and SDR monitors can be on air side by side and both look right. Switching HDR on or off in Windows during a session is noticed and the capture comes back in the new format on its own. When the output is SDR every HDR source is always brought down to SDR.
• Record in HDR10. OUTPUTS › COLOUR: HDR makes the canvas 16-bit float and writes PQ Rec.2020 10-bit to the file. This uses HEVC Main10 with full HDR10 signalling and mastering metadata. Preview, NDI and screenshots get a tone-mapped SDR picture of the same frame. Streams deliberately stay SDR. The block shows the HDR/SDR state, white level and peak of every display.
• Auto exposure, the way a game engine does it. Metering uses a 64-bucket log-luminance histogram computed entirely on the GPU, over every pixel in linear light. Its metering mask offers full frame, centre-weighted or spot. The darkest and brightest percentiles are discarded. Adaptation works in stops, with separate speeds up and down. The correction is an exposure multiply rather than a gamma bend, so darkening really recovers blown highlights. An EV compensation dial and hard min/max limits were added; the per-frame CPU copy and measurement of the frame are gone.
• Fixed: the deck opened as a blank white window (settings window black, projector black), on machines with virtual-display drivers. Parsec Virtual Display, spacedesk and similar tools register an "indirect display" adapter listed by Windows under the real graphics card's name and VRAM. A two-card machine showed four identical cards in PERFORMANCE. Picking one of those nodes as card #0 made the app render every frame and present it successfully. Yet nothing it drew reached the screen. Every adapter is classified through the kernel's adapter type. Nodes without render support are skipped at startup and listed as non-pickable in PERFORMANCE. Card numbering counts real cards only. Each card shows which monitors it drives. The log names every card's LUID and displays. In the first seconds, the deck reports presentation status and the mean colour of the frame handed over. A blank window can be diagnosed from the log alone.
• Fixed: drop-down lists, context menus and tooltips opened away from the cursor. The desktop guard for multi-monitor walls with gaps treated anything smaller than 120×32 px as lost. Every popup starts as a 16-px placeholder on its first frame, while content is still being measured. The NDI source list, menus and tips landed in the middle of the deck's monitor instead. Popups are left to the UI toolkit, which already clamps them to the monitor under the cursor. The threshold scales with the window. The start-up placement self-check gained two cases.
• Settings in the cloud, by hand, both ways. LICENSE › CLOUD SETTINGS SYNC has two buttons. Upload to cloud replaces the cloud copy with this computer's settings. Download from cloud applies the cloud copy to this computer after confirmation, keeping every local secret. API keys and stream keys are never copied. The tab shows the revision, date and app version of the cloud copy. Nothing moves on its own any more. The previous behaviour is an opt-in Automatic checkbox: upload after each settled burst of edits, restore a fresh install at start.
• Fixed: an accessibility client could freeze the app for 20 seconds. A screen reader or automation tool could request a UI tree from the deck, floated settings panel or projector (WM_GETOBJECT). The system's UI Automation server answered by parking the interface thread until the watchdog gave up. The windows are entirely custom-drawn, so the request is now answered in place with no provider and no inter-process call.
• Windows that heal their own presentation. The Windows compositor could stop taking deck or projector frames on multi-monitor rigs spanning two cards. The app forces resynchronisation after about 1.5 s and rebuilds the swap chain if needed, instead of showing a blank window indefinitely. Projector windows resize their swap chain on DPI or mode changes. Double-clicking a full-screen projector makes it a normal resizable window (two thirds of the monitor, centred), and back.

Withdrawn: superseded by 0.0.77a. No longer distributed.

Every colour panel gains brightness, saturation and vibrance. An extra camera keeps its own grade while on air. One-click screenshots cover any preview and the programme.

• Brightness, saturation, vibrance, per source and on the output. CAMERA PICTURE gains three sliders next to contrast for the main camera and each extra camera, picked at the top. Output COLOR GRADE gains them too. Brightness is a plain level before the contrast pivot. Saturation sits around luma (0 = mono). Vibrance lifts less-saturated areas first, so skin comes alive without neon. On the output, vibrance moved from the auto-exposure block into COLOR GRADE, where the rest of the colour controls are.
• Fixed: an extra camera on air was graded with the main camera's curves. When the director cut to "cam 2", the programme used the MAIN camera's grade and ignored its own. Grading that camera changed nothing while it was up. Whichever camera is on air now gets its own grade and input LUT, in the same single pass the tiles and the inset already used.
• Fixed: a camera transition showed the same picture in both slots. During a slide, whip or rotate away from the main camera, the second half drew the outgoing slot with the incoming camera. Cutting back drew the incoming slot with the outgoing camera until the midpoint. The cut stepped, shifting the picture in the inset and on its backdrop. Each slot now shows its own camera for the whole move. Cuts between cameras of different resolutions no longer rebuild the backdrop blur chain on the cut frame. The retiring size is parked and reused on the next cut. The cut itself no longer hitches.
• Smoother frames, lower camera latency. A live-show profiler found several causes of lost frames. Without a Stream Deck, reconnect walked and opened every HID interface on the render thread: 64 ms every 2 s. It now runs on a worker. The 2-second settings autosave serialised the whole config on the render thread, even without changes. It now uses a hash walk, well under a millisecond. On-screen intelligence scanned the app's own windows through UI Automation, answered by the render thread. This took up to 1.2 s of message servicing every 30 s; own windows are now skipped. While open, device pickers synchronously enumerated cameras, capture cards and audio endpoints every 3 s, causing stalls up to 2.4 s. Enumeration now runs on a worker. Camera frames reached the GPU through a cached memcpy into write-combined memory: 1.9 ms a frame for two cameras. They now use streaming stores. The webcam collector allocated and zero-filled a fresh 8 MB buffer per 1080p frame, then walked it again to force alpha. It now uses one SIMD pass over a recycled buffer on an MMCSS "Capture" thread. A backlog guard skips stale samples, preventing the device queue from becoming visible delay. The cursor tracker was left exactly as it was. Any frame exceeding its budget is logged with the longest stage. This is how the Stream Deck probe was caught.
• Screenshots. A camera button on every source preview saves a PNG of exactly that view at the source's resolution. It includes colour grade and tracker overlay when each is ticked. The same button sits in the programme monitors' top-right corner (16:9 and 9:16). It saves at canvas resolution, after the whole output grade. Files land next to your recordings, in screenshots. A dry camera-shutter click confirms the shot; it has its own switch in DIRECTOR › INPUT SOUNDS.

Withdrawn: superseded by 0.0.76a, which fixes the deck opening as a blank window on machines with virtual-display drivers. No longer distributed.

Per-camera settings now follow the camera you picked, and the source previews got two switches you can actually see.

• Fixed: per-camera settings could land on the wrong camera. The live extra-camera list skips slots with no device or devices that fail to open. So "cam 2" could become first and inherit "cam 1" settings. Colour grade, input LUT, eyeline enrolment and STUDIO seat microphone are keyed by the slot number in MORE INPUTS. Unplugging one camera no longer hands its look to another.
• Source previews: two switches, in the view. "trackers" shows the followed face and crop actually taken from the source. "colour grade" shows the camera through its grade and input LUT instead of the raw feed. You can match cameras while looking at what goes on air. Both were only in a right-click menu before; both are remembered per view.
• Where the per-source grade lives. SETTINGS › SOURCES › CAMERA PICTURE has a source picker at the top. The main camera and each extra camera keep their own curves and .cube.

Withdrawn: superseded by 0.0.75a, which grades an extra camera with its own curves while it is on air. No longer distributed.

Three rooms in one app: SCREENCAST, STUDIO and EDIT. STUDIO cuts a multi-camera talk show by speech; EDIT holds the recordings library and editor. There is a first-run wizard, rebuilt music player and fixes for the full-screen hang.

• Fixed: the app could crash or hang when going full screen. Crash reports showed the projector reading an empty monitor table as if it had entries on a single-monitor machine. It went down with the whole window. It now checks first. If the full-screen projector takes the only monitor also captured by the app, it hides itself from capture. The picture cannot feed back into itself. Two more null-pointer guards from the same reports (chat overlay, the window backend).
• SCREENCAST / STUDIO / EDIT. One switch in the menu bar changes what the app is for. Each mode keeps its own workspace (the strips you closed, where you docked them), so the cutting room does not undo the cockpit. EDIT is refused while you are on air: the renderer is busy.
• STUDIO: a studio director for video podcasts. Every camera is a seat. The director cuts to the speaker by TV rules. A "mm-hm" never steals the shot; nobody is cut faster than a beat. Two people talking over each other get a two-shot. Silence and long monologues go to the room. A listener's reaction is shown, then it comes back. Speech is read from each seat's own lips (the camera sees the mouth move) fused with the room mic. A temperament slider, and click any seat to hold it, WIDE to hold the room, AUTO to hand it back.
• STUDIO tracks every camera at once. Each seat runs its own face track with the screencast camera's deadzone, spring and snap-on-cut. All cameras frame their faces simultaneously, not just the one on air. A tile therefore arrives already composed instead of drifting into place.
• STUDIO MOSAIC: flat, fitted pictures instead of floating screens. The wide shot is called MOSAIC and built by the canvas. 16:9 uses columns: two side by side, three columns, four in a square, five as three over two. 9:16 uses rows. No frames, no halos, no tilt. Cuts between two full-frame shots are instant. Entering or leaving a mosaic slides, with tiles parking off the nearest edge and returning from the side they left. The gap and the slide tempo are yours to set.
• STUDIO: a microphone per seat. Each seat can use a wireless receiver, capture card audio, a second interface input or "automatic". Automatic pairs the camera with the audio endpoint sharing its name. That sound joins the voice bus, appearing in the voice stem and on the stream. It becomes the seat's speech detector, better than reading lips. Per-seat trim, and a pick that would open the main microphone twice is refused.
• STUDIO has its own settings, and its own pad. SETTINGS gains a STUDIO page with cutting rules, mosaic and seats. Pages for the other show, including screencast autopilot and browser-guest room, are hidden there. The seat panel is now the same grid of big tiles as the screencast MODES pad, with PAUSE, PANIC and CAM alongside.
• Colour correction and an input LUT for EVERY source. The camera picture page has a source picker: the main camera and each extra camera keep their own curves and their own .cube. The correction is computed once per frame and follows that camera through backdrop, inset and studio tiles. A log capture card, warm webcam and phone can match one look.
• EDIT: a recordings library you can actually read. Takes are grouped by day with a poster frame and capability badges. These show RE-CUT, cameras, screens, audio stems, transcript, a sitting in progress and a rendered film. One row of verbs: Edit, Play, Analyse, Transcribe, Folder, Delete.
• Transcribe every audio channel. Microphone, system sound (remote guest or video) and full mix are transcribed separately, each to its own subtitle file. The microphone also becomes the take's captions. Priced before anything is sent, like every paid action in the app. Captions fall back to the live subtitles the session already wrote for free.
• The editor, simplified. A three-step strip, LOAD, CUT, FILM, and the deep controls (automation curves, the AI planner, the music sources, the length target) behind one "advanced" switch. Multi-camera takes get a camera column in the cut list: pick the camera per piece; the player and the render both follow it.
• First-run wizard. Four short steps, what will you make, the picture, the sound, where recordings go, that write the same settings the SETTINGS window edits. Help › Setup wizard runs it again.
• UI scale. View › UI scale, 80–200%, rebuilds the fonts and the spacing live.
• Music player rebuilt on two decks. The next track decodes ahead on its own deck for seamless skipping, seeking and automatic changes at track end. An equal-power crossfade (the new "track crossfade", 0.1–12 s) starts exactly at the file's last sample. The outgoing track runs while fading. The incoming one begins where you asked. No gaps, no clicks, no repeated intros. Verified with synthetic tones: no level dips at any join.
• Mute camera rethought. Muting freezes and blurs the last frame while the scene re-flows without the camera. PiP goes quiet; the camera's ISO track records a black slate. A sound effect marks the change and the CAM tile lights up. The state survives a restart.
• The cursor moves like a hand. The pointer is sampled at 250 Hz with a least-squares velocity estimate, smoothed by a One Euro filter and driven by closed-form springs. It avoids lag during fast moves and wobble at rest, at any frame rate.
• Restart button. Restart-only settings now share one table, so RESTART appears whenever one changes. The button is also visible in the SETTINGS window, not only the deck.
• CINEMATIC hides the pointer only on a film that is playing. The drawn cursor now disappears only while the video actually plays and the pointer sits on the player. A paused player, or CINEMA forced on an ordinary window or app, keeps normal pointer rules. What you point at stays visible.
• Sound effects stay out of TALK, AMBIENT and the film. The whole SFX bus is muted in TALK, AMBIENT and while video plays in CINEMATIC. This covers clicks, key ticks, whooshes, cursor wind and drag. A short ramp prevents clicks. The recording's SFX stem falls silent with it.

Withdrawn: superseded by 0.0.74a, which fixes per-camera settings landing on the wrong source. No longer distributed.

The app can no longer refuse to open because of one of its own settings. New: a welcome screen that shows what it is loading, touch that leaves your mouse alone, and per-format controls for 16:9 and 9:16.

• Fixed: the app could refuse to start, with no way back in. If a canvas failed to initialise, the app stopped, reporting that no format was enabled. Such a failure can happen transiently, for instance right after the app was force-killed and the graphics card has not released it yet. Both formats were enabled; the message pointed at the wrong thing. Worse, the place to change those settings is inside the app that would not open. It now falls back to one modest canvas and says plainly what happened, so there is always a way in.
• Touch no longer throws your mouse pointer across the screen. Tapping a mode tile used to teleport the Windows cursor to that tile. Windows turns an unhandled touch into a synthetic mouse click, and that click has to happen somewhere. Touch and pen are now handled directly, so the pointer stays where you left it. Useful if you hold a mouse in one hand and tap with the other.
• Each format has its own controls. OUTPUTS › CANVASES: 16:9 and 9:16 each get their own size, frame rate, graphics card and on/off switch. These existed only in the config file before. Preview thumbnails have their own refresh rate too. It never affects what you record.
• Previews only cost something when you can see them. A preview hidden behind a tab was still being rendered in full, sixty times a second. That is what made the desktop and the mouse feel heavy. Measured on a five-monitor workstation with the 9:16 preview ticked but not visible: from 37 to 60 fps. Dropped frames went from 82 to 1 per 120.
• A welcome screen that tells you what it is doing. The app opens with a card showing the actual loading stage (vision models, audio, director, sources) over a looping backdrop. It fades out on the first real frame. Also available any time from Help › About.
• Crash reports now say which build crashed. They were stamped with the version doing the uploading, so a month-old crash arrived labelled as a fault in the newest release. Dumps also went to a folder that depended on where the app was launched from. On an installed copy that folder is often not writable, so the crash we most wanted to see left nothing behind. Both fixed, and one hang no longer files two support tickets.
• Camera work: hand close-ups now frame the hand (the tracker only knew the fingertip, so a close-up put the hand half out of shot). They choose between both hands and one from the geometry rather than at random. Cuts land on their framing instead of gliding into it. TALK can be presenter-only and centred, with the window sliding out of frame. A video wallpaper plays when there is no camera.
• NDI: the receive bandwidth was set to a value the runtime does not define.

Withdrawn: superseded by 0.0.73a, which fixes a full-screen hang this build could hit. No longer distributed.

Critical fix. Install this one. On 0.0.70a the app could freeze in a way that took the whole desktop down with it. Please update.

• Fixed: a freeze that locked up the entire machine. Windows restarts its taskbar on its own, after a display change, an update, or an explorer crash. The restart broadcasts a message to every running program, telling it to put its tray icon back. We put ours back on the spot, while still handling that message. Putting an icon in the tray is a call into the taskbar that waits for an answer, with no time limit. If the taskbar was not answering yet, which is exactly the case while it is starting, we waited forever. A program that stops answering messages then blocks everything else that speaks to it, so the desktop went down too. Talking to the taskbar now happens on its own thread that nothing waits for. It retries while the shell is still coming up instead of silently losing the icon. The message handler returns immediately.
• Fixed: closing a panel could put the layout into a loop. Reclaiming the space left by a closed panel was being re-checked on every single frame, which could keep rebuilding the layout instead of settling. It now happens once, when a panel closes.

Withdrawn: superseded by 0.0.72a, which fixes a start-up failure this build could hit. No longer distributed.

Record a conversation with other people and everyone ends up holding everyone’s source files. Also in this release: a long list of things in the cockpit that looked finished and were not.

• A shared take across machines. Everyone in a room ticks “ready”, the host arms it, and recording starts on every machine at the same instant. Two computers disagree about “now” by hundreds of milliseconds, and a “start” message arrives whenever the network feels like it. That is why the room server hands out the clock. Measured across two live instances: they fired in the same millisecond.
• Afterwards everyone gets the real files. What travels live is sized for the network and is not editing material. When the take ends, each machine uploads what it recorded (full bitrate camera, screens, audio stems, the session data). It also pulls down what the others recorded, into the same take folder. Sent in chunks that resume instead of corrupting the file if the connection drops, and verified byte-for-byte.
• Fixed: the camera framing slid into place after every cut. A cut handed the camera a composition solved for the shot that had just ended. The composition then eased into position over the best part of a second. That was visible as the picture “riding” in the backdrop and in the inset. The head’s tracked size was the half nobody had noticed. It kept easing while the position teleported, so a cut landed on the right spot and then zoomed for half a second. Both are now correct on the cut’s first frame, and ordinary tracking still glides.
• Fixed: two audio meters never moved. The UI sound set and the room guests were only ever measured on the path that runs when transmission delay is switched on. So during ordinary work those bars sat dead. A meter that stands still is worse than no meter: it says “silence” where there is signal. While checking the rest, two more were reading only the left channel, so material louder on the right under-read or vanished.
• The mixer reads like a mixer. Every bus is a full channel strip: fader, peak meter with hold and a clip latch, gain-reduction bar, mute and solo. The channel’s own settings sit behind its name. Gate, compressor, ducking and the limiter are visible as they work, which is the part a level bar can never show. Columns are now named at the top as well as the bottom, so it is obvious which fader belongs to which channel.
• Kick. Requested through the in-app support form. It sits with the other platforms. Each of them now folds away on its own and says whether it is keyed, ticked and on air even while folded.
• A cockpit that uses its space. Mode tiles fill their panel instead of leaving half of it empty. Transport buttons share out the width and carry their names when there is room, falling back to icons when there is not. The start / pause / auto buttons form one block at the left edge. Closing a panel no longer leaves a grey hole where it used to be: its neighbour takes the space back. Dragging a splitter now squeezes the panel next to it instead of stopping dead.
• The editor got the rest of what it needed: pieces can be split, moved, copied and trimmed. The player can run the CUT rather than the raw take, so joins and speed ramps are audible and visible before rendering. The recordings list has thumbnails, sorting and a filter.

Withdrawn: this build could freeze the whole machine (fixed in 0.0.71a) and is no longer distributed.

The recording you edit is now the whole session, and the editor plays your cut rather than the raw take. This release is mostly repair, several things that looked finished were quietly not.

• Fixed: long multi-monitor recordings were silently losing part of the session. One encoder chip has a finite pixel rate, and five 4K screens plus cameras could ask for roughly twice what it delivers. Nothing pushed back: the shortfall accumulated until it looked like the machine had been asleep, and that stretch was dropped from the track. A 232-second session came back as a 180-second file, internally consistent, in sync, and missing a fifth of what happened. Encoder load is budgeted before recording starts. Separate tracks use a rate the hardware can hold: 30 fps by default, lower only when the machine cannot keep up. The app reports a lower rate.
• The separate tracks now share one clock. Each track used to anchor its timeline on its own first frame. Files looked right individually but could not be laid against each other. Frame N of the camera is now the same moment as frame N of a screen, which is the entire premise of cutting between them.
• Fixed: sources stuttered while recording. A recorded monitor rebuilt a full 4K mip chain on every captured frame, on every GPU. Its consumer only reads the full-size image.
• Fixed: the editor could not actually edit. The session recording and edit list kept time differently. A render produced one frozen frame for the whole film; the player showed no picture. Both now run on the take's own timeline, including recordings made by earlier builds.
• The player plays the cut. A switch puts transport on film time: only retained pieces, in order, at their own speed. Dropped material is skipped at joins. Original sound is muted under sped-up stretches, exactly as the render mutes it. Joins and speed ramps can be judged before rendering anything.
• A timeline you can work on. It always showed the whole recording, about seven seconds per pixel on a two-hour take. It zooms around the cursor and has a ruler in hours or hundredths of a second. Drag a piece's edges to trim it. Pieces can be split at the playhead, moved, copied and pasted whole, speed, shot type and zoom keys included.
• The recordings library gained thumbnails, sorting and a filter. Both library and editor have View menu entries, so closing the editor is no longer a one-way door.
• The editing assistant can read the session. One paid pass transcribes speech and uses a vision model to describe frames at notable moments, storing both with the recording. The plan comes from what was spoken and seen rather than from a curve. Re-planning afterwards is free. The price is shown before anything is spent.
• Editing is armed by default. Recording edit data used to require ticking a box before pressing record. Missing it meant the session could never be re-directed. It costs about 77 MB an hour next to gigabytes of video.
• Multi-monitor and virtual sources are editable. The film follows the screen on air instead of always using the first. NDI feeds, framed windows and regions are editable, with their original framing rather than the padded surface.
• Target length, burnt-in captions and the soundtrack choice now do something. Target length is reached by speeding up only stretches without speech. If that would require cutting into speech, the render says so instead of speeding up a voice.
• Fixed: a rendered film carried no watermark while the live picture did, so the same material could be exported without it. The licence rule is now the same for both.
• Fixed: closing the app while an AI job was running killed the process. Also fixed: playback overwrote your own director settings and never gave them back, and a closed editor's automation kept driving the live director.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

The films get made. A recorded session runs back through the broadcast engine. An edit can change the directing, not just where cuts fall.

• Re-direct a session instead of trimming it. The recorded director inputs, settings and theme go back into the same compositor, with the original sources decoded on the GPU. Different shot types or tighter zooms on chosen keyframes are re-rendered at full quality. A vertical cut is framed afresh from originals rather than cropped from the finished file.
• The same edit gives the same film. Verified rather than assumed: every frame of two separate renders folds into an identical digest. The timestep comes from the output frame rate, not the wall clock, which is something a live pass could never promise. A byte-identical file is available too, as a mode, since it costs some quality per bit.
• The soundtrack is rebuilt from the separate stems, microphone, music, system, effects, so an edit can disagree with the balance chosen live. Joins get an equal-power crossfade. A hard splice between pieces of the same room tone clicks, telling the viewer where the cut was.
• Timelapse that cannot speed up speech. A speed request is split around speech, which plays at normal rate. The rule holds by construction, so no edit accidentally produces a chipmunk voice. In fast parts, original sound drops out and music keeps its tempo. A synthesized typing texture is rebuilt from recorded typing density.
• Privacy masks are restored unconditionally. Source tracks hold the raw desktop. The recording carries what was hidden and the render restores it. No setting can lift it. A recording made before this release is treated as unknown, and says so, rather than as nothing to hide.
• Renders wait for the machine to be free. Nothing renders while recording or streaming, not slower, nothing, because a paused render still holds encoder sessions and video memory. You are notified when a film is done.
• An editing assistant can propose the cut from the free local analysis, with the cost shown before anything is spent. What it returns is checked against the recording and trimmed where it disagrees. It has no way to lift a privacy mask or to speed up a voice.
• Fixed: rendered films were losing frames and had no sound. Both were invisible in the logs, which counted frames submitted rather than frames written.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

The editor opens. You can now pick the moments that matter out of a whole session, without watching it back. Turning those choices into a finished film is the next step; this release is about choosing.

• See where anything happened, at a glance. The session is drawn as five heat strips, overall interest, voice, activity, face, and your own markers, so a two-hour recording becomes readable in seconds. Drag across any of them to select a stretch; the strip also shows which app was in front at that moment.
• Pick, or let it pick. Add ranges by hand, or press Auto-pick best for a starting point built from the highest-scoring moments. Your markers outrank everything the analysis computed, a moment you flagged during the take always wins.
• Choices are saved as an edit decision list, not as a cut of the video. It records source ranges and direction changes: a different shot type, monitor or tighter zoom. The film will be re-rendered from original sources rather than trimmed from the old one.
• Open any editable take straight from the recordings library.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

First groundwork for the native editor. A take can now be recorded as a project rather than a finished video. The app writes down everything it knows while it records, so a session can later be re-directed offline instead of merely trimmed. The editor itself comes next; what ships here is the recording, the library and the analysis.

• “I want to edit after this session”. One switch in OUTPUTS. It states what gets turned on and roughly what it costs in disk, then turns it all on with one click. Off by default: a normal recording is unchanged.
• The session tape. Every frame, the app stores what it knew: cursor path, clicks and scrolls, window titles and the text caret. It also stores face and gaze, voice activity, music energy and cue points, and the shot the director chose. Plus a full state snapshot every two seconds. With it, an editor can later jump to any moment and see exactly the frame that would be there.
• Mark the good moments while you record. Two bindable keys (also on a Stream Deck): one says “this was good”, the other “never show this”. It is the one editing signal no analysis can infer, because it carries intent.
• A recordings library. Every take with its date, length, size and resolution, read from a small manifest rather than by opening video files, so it appears instantly. Deleting is explicit and confirmed with the size; nothing is ever removed on its own.
• Find where anything happened. One click builds a local map of the session. It shows a per-second interest curve, which app was in front and when, where the dead air is, and your markers. It also shows the character of the session (tutorial, conversation, research, quiet work). Runs offline and costs nothing: no AI is involved.
• Smaller ISO files where it matters. Editing tracks now use AV1 on GPUs that have it (roughly a third smaller than HEVC at the same quality). The delivered recording keeps HEVC for compatibility.
• A warning that used to be silent: if a GPU runs out of concurrent encoder sessions, some ISO tracks never start. That loss cannot be recovered afterwards. The status bar now says so during the take.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Setup AI tells you what went wrong. Every network failure used to arrive as “HTTP 0”, a number that names nothing. The app always knew the real reason, so it says it now: no DNS, connection refused, timed out, or a proxy / VPN / antivirus intercepting HTTPS. If the AI master switch is off, the Setup AI window says so up front. It offers a button to turn it back on, instead of letting the request fail silently. Genuine network blips are retried once, automatically.
• Scrolling has its own sound. It is a soft breath of air rather than the tick a click makes. It rises in pitch while you keep scrolling and settles when you stop. A rate limit plus per-notch pitch variation stop a fast flick of the wheel from turning into a rattle. Eight dials: volume, tone, length, pitch variation, rise, max rate and stereo width.
• Fixed: the shot opened up too late in Focus mode. A planned zoom-out was being treated as an ordinary release. As a result, the zoom-hold timer froze the frame for exactly the head start the delay buffer had just bought. An anticipatory widening now runs on its own clock. The frame is already wide by the time your hand moves.
• Focus no longer drifts back to the mouse while you type. It used to give up after a few quiet seconds and swing the frame onto a motionless pointer. A pause mid-sentence was enough. Following the text caret is now released by the mouse moving, and by nothing else, so the view moves the moment you move the cursor.
• The cursor no longer blinks between shapes. Windows swaps the pointer handle constantly at link and text boundaries; the drawn glyph was following every single flicker. It now waits for a type to settle, and waits longer the more the pointer has been flapping. Then it morphs one shape into the other instead of swapping them.
• Hide the native cursor over the fullscreen preview. It is now an option rather than fixed behaviour. The projected image draws the app's cursor effects, so the Windows arrow on top doubled the pointer. This applies with click-through off. With it on, the pointer belongs to the app underneath.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Fixed: edits made from the settings search did nothing. Rows in the raw config sections wrote the value to disk but never told the subsystem, so nothing changed on screen. For sections backed by a settings struct, the next save wrote the old value straight back over yours. Those rows now go through the same live-apply path the Setup AI uses, which also gained the theme keys it was missing.
• Config rows are real controls now. Values with a documented set of options become a dropdown instead of a text box you had to guess at. AI model, voice and TTS keys get a Choose… button with a filterable list of what the app has loaded. Folder and file keys get Browse…; r,g,b keys get a colour swatch.
• A proper menu bar. File, Edit, View, Settings, Help. It owns no settings of its own: every item drives something that already exists, and transport entries show the shortcut you bound in CONTROL. View switches each of the seven layout sections on and off, and remembers your choice.
• The cursor is choreographed, not chased. Behind Transmission Delay the drawn pointer now glides from anchor to anchor. Clicks, button releases and scroll notches keep their exact timing; the path between them is designed. No more freezing in place and then darting across; the curve has no moment where you can see the movement begin. Four modes (raw, smoothed, animated, tutorial) plus seven dials.
• The shot opens up before you move. Anticipation now scales with the buffer you hold instead of a fixed three quarters of a second. How far it pulls back scales with how far the pointer is about to travel. The frame also follows the pointer you see rather than the raw mouse.
• Hiding the app from capture now means the whole app. That is the deck, the SETTINGS window and every panel you floated out. Previously only the deck vanished and SETTINGS stayed on the recording.
• A RESTART button appears when, and only when, one is needed. Settings the app reads once at launch (render fps, recording resolution, canvas GPU, vision engine) raise a small yellow badge in the status bar. It lists exactly what is waiting to be applied. Click it to restart; it stays disabled while you are recording or live.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Search in your own language. Type in Polish, German, Spanish or French and the search still finds the setting. Accents and case are folded away (“głośnośk” behaves exactly like “glosnosc”). A 43-group vocabulary maps every term onto the English catalogue.
• It finds everything now, including CONTROL. Keyboard shortcuts and MIDI-mappable targets live in config arrays the old scan could not express, so the whole section was invisible. Every action and every knob target is searchable, shows what it is bound to, and links straight to the CONTROL tab.
• Better answers for longer questions. A strict “every word must match” rule used to hide most useful results. Rows matching all your words come first, then the progressively weaker sets, each clearly labelled. Matching a whole word now outranks matching letters buried inside one.
• Setup AI can search too. It is no longer handed the entire catalogue on every request. It now runs 1–5 targeted searches first (always in English, whatever language you wrote in). Then it plans from just the settings those return. That is cheaper, and far more likely to touch the right knob.
• Richer hints everywhere. Descriptions across the parameter map now name both words for the same thing (cursor / mouse, screen / monitor, camera / webcam, subtitles / captions, shortcut / hotkey). That way, searching either one works.
• Fixed: the PERFORMANCE toggles did not survive a restart. They were saved to a different config section than the one they are read from.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• The workspace scales proportionally now. Resizing the window used to push the entire change into the preview while every other section kept its exact pixel size. As a result, the layout fell apart a little more with each resize. All sections now keep their proportions, and the sizes you set by hand are preserved as ratios.
• Splitters now shrink their neighbour. Dragging a section edge stopped early because a large minimum was reserved on both sides of every splitter. A section can now be pushed nearly closed (its content clips), and the separator is easier to grab.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Fixed: the app got heavier the longer it ran. Window-gallery captures were never released. The 2-second scan kept adopting new windows and every one of them stayed alive, on every GPU. They are now reclaimed like screen captures, which is the single biggest cause of the creeping lag.
• Half the capture sessions at launch. A session was opened for every monitor on every graphics card, even one whose canvas was disabled. Cards now open captures only when they compose.
• Detection no longer fights your game. The face / hand detectors ran on every single frame: 60 Hz of GPU inference competing with whatever else you were running. They now run at a configurable cadence (12 Hz by default), which looks identical and gives the time back.
• GAME mode is now a no-lag mode. It renders the captured screen as it is: the whole post chain (exposure, vibrance, sharpen, LUT, grade, chromatic aberration, vignette, scanlines, dither, bloom) is bypassed. It also sheds the window gallery, off-air captures and supportive-camera detection, and asks the driver to schedule our GPU work below the game's.
• New PERFORMANCE tab. Live resource readout (capture sessions, capture VRAM, delay buffer, free memory on the tightest card). Plus every policy switch above, each with an explanation of what it costs.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• The shot widens before your hand moves. With a delay buffer running, the zoom-out now follows a trapezoid. It eases open during the pause that precedes a move, holds wide through the travel, and returns only after the pointer has landed. It is a glide instead of a reaction.
• Raw config keys became real settings. Every key found by the search now gets a typed control (checkbox, integer slider, float slider or text field). A sensible range is inferred from the key's unit convention (dB, ms, s, Hz, MB, kbps, px). The unit is printed in the value. They are tinted amber so they never read as a curated setting.
• Rich hints everywhere. Those raw keys now show the full description from the app's own parameter documentation instead of a bare config path. It covers what the setting does, how it relates to others, its type, unit and range.
• Fixed: clipped mode buttons. The square tile grid reserved too little room for the controls beneath it, so tiles were cut off. It now measures the real row height and never overflows its section. The quick sliders below adapt their width instead of running off the edge.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Fixed: settings search was read-only. Dragging a slider in the search results snapped straight back. The value was re-read from the app every frame, so nothing could be changed. Search results are now fully editable, and three filters (no presets, described only, group) keep the list meaningful.
• The pointer can no longer teleport. Every move rides an ease curve (fastest mid-flight, braking into the target). Wandering is paced calmly instead of darting. A continuity guard with a hard per-frame ceiling makes a jump mathematically impossible. The screen now follows that corrected pointer, so the framing stopped reacting to raw mouse jitter nobody sees.
• INTRO / OUTRO respect the transmission delay. With a delay running, the outro used to slam over content from seconds ago and cut the end off your take. Stinger requests now fire on the content clock.
• The delay slider tells the truth. A green band marks the range your GPU can sustain right now (the rest is dimmed). A live readout shows running / achievable / buffered / pool. You can still ask for the full 10 s.
• Smoother mode changes. Switching from WORKING to HYBRID no longer jerks. The cut used to inherit the previous shot's framing momentum while the zoom mapping changed underneath it.
• Mode buttons are a touchpad. Square tiles, auto-arranged into the best grid for the space they are given.
• PiP: control over the inner canvas. New dial splits the frame's beat pulse from the picture inside it. Hold the image rock steady while the frame breathes, or let it over-react.
• The AI co-host stops mixing up old and new. Memories carry exact timestamps and reach the prompt with a readable age (“14 min temu”). They expire after a configurable window and arrive as one ranked block instead of scattered fragments.
• Also: the 9:16 dials fold into their own section, and the LUT picker now reports exactly why it failed instead of doing nothing.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• The pointer is a planned animation now. Clicks are the keyframes: the drawn cursor holds on the last click (or wherever your hand parked). Then it makes one deliberate, eased move timed to land exactly on the next click at its exact moment. The raw jitter in between is ignored, and the whole plan cross-fades back to realtime as the transmission delay shrinks. Zooms couple to the same plan: wide during the journey, punch on arrival. Held buttons still track the real mouse pixel-for-pixel.
• Face-aimed camera pick, zero setup. The Hybrid backdrop now switches to the camera you are facing. Every feed already runs its own face tracker, so the app ranks cameras by how straight-on your face is. It cuts to the clear winner while you speak. The manual seat-position sliders are gone: nothing to configure.
• Search that now finds things. The settings search now covers every config key in the app (hundreds), not just the curated registry. It understands Polish/English synonyms (mouse = cursor = mysz = kursor, dzwiek = audio, pion = 9:16). It lays results out wide, with sliders that stretch with the window.
• One more 9:16 dial. "Hybrid split shift" moves the face/screen divide of the vertical Hybrid shot. The portrait canvas now reacts visibly in the shot you use most.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Search every setting. A quick-search box now sits above the SETTINGS tabs. Start typing and the tab stack becomes a ranked list of every matching setting in the whole app. Every row is directly editable in place (sliders with their real ranges, checkboxes, enum pickers). Labels match at full weight, the long per-setting hints at a lower one, and results settle with a natural typing rhythm. Hover any row for its full description and config path.
• MIDI knobs behave. Fixed the temperament mapping (it covered only half the slider and snapped the value down whenever AUTO had pushed it above 1.0). Added soft-takeover for every absolute CC. A knob engages only once it reaches the current value, so touching it never makes the parameter jump. Every numeric setting in the app is now MIDI-mappable automatically, with its real range transferred onto 0–127 from one source of truth.
• Seven more dedicated 9:16 dials. Focus presenter band, WORK/FOCUS side paddings, Ambient card size and gap, and both edges of the protected face zone. All previously hard-wired, all with individual hints, all searchable and MIDI-mappable.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• The pointer moves like an animator drew it. Smoothing is now driven by the montage temperament (calm show = long cinematic glides, hyper = tight follow) with velocity only as a mild trim. Between two nearby clicks the pointer travels a designed, eased point-to-point curve and lands exactly on each press. The moment any mouse button is held, it sticks rigidly to the real mouse, so drags and selections stay pixel-true.
• The frame follows what you type. Park the mouse for two seconds while a text caret is alive and the attention frame moves to the caret. This works on both the 16:9 and 9:16 canvases. With the transmission delay active it even flies back to the pointer just before your next click appears.
• Dedicated 9:16 framing controls. New VERTICAL section in DIRECTOR. It holds zoom boost (how much harder the narrow canvas punches), presenter band height for the portrait WORK shot, and the vertical Hybrid zoom slope. The portrait canvas now has its own dials instead of hard-wired multipliers.
• Ambient gets real depth. The window gallery reacts to your head position (pure translation, no rotation): deeper cards travel farther, so the desktop stops hanging flat in space. Strength slider included, 0 turns it off.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Transmission Delay now reaches its full 10 seconds, even at 4K. The buffer keeps frames on the GPU in NV12, 3.5× lighter with no visible loss; every encoder ships 4:2:0 anyway. The VRAM budget sizes itself automatically from the free memory of your tightest GPU. Together that turns “0.4 s effective” on a 4K rig into the whole slider. Unused captures release memory: deselected monitors and screens of a disabled canvas free their sessions after a few seconds. They return instantly when needed.
• The pointer stops darting around. With delay active, the drawn cursor follows a zero-lag smoothed path from its past AND future, possible only with a time buffer. It tightens during fast deliberate sweeps and lands exactly on every upcoming click. The director also zooms out early enough, and far enough, to keep the cursor’s whole upcoming path inside the frame.
• Hybrid can follow your eyeline. While you speak, the backdrop can switch to the camera your head points at. Set each camera’s seat position with a left-front-right slider. With the delay active the cut lands as your head settles, before the first word goes out.
• ISO tracks: no more MOV. The camera and every monitor record crash-safe and deliver as clean faststart MP4s (stream copy, zero quality loss). Also fixed: a motionless monitor recorded nothing because its only frame arrived before anyone was listening. Screens whose capture never matches the encoder now report it in the log instead of failing silently.
• Cockpit polish. FPS readouts show two decimals and grade against YOUR configured frame rate. Workspace splitters no longer jam when SETTINGS is docked. Docked SETTINGS survives restarts. The “fullscreen when window >” slider shows real percentages.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Transmission Delay: a director that knows what happens next. A new 0–10 s slider (SETTINGS › DIRECTOR) delays your whole programme: picture, sound, streams, recording, subtitles. Meanwhile, the director keeps watching the live signals. Working that far ahead, it pulls the shot back before a big mouse move happens and punches in before you click. It cuts to a fresh shot just before you switch windows and ducks the music before the first word. Viewers see framing that looks like someone read your mind. At 0 s everything stays exactly as it was: the real-time path is untouched, bit for bit, with no added cost. Frames are buffered on the GPU with no extra copies. Privacy masks travel with the pixels they belong to, and the slider locks while outputs are running.
• Fullscreen preview you can work through. The fullscreen projector now passes mouse and keyboard to whatever is underneath (on by default). So the clean programme can cover a monitor while you keep using Windows normally. Leave it via the panel's CLOSE PREVIEW, a bound FULLSCREEN hotkey, or the top-right hot corner. Park the pointer there and a × chip appears.
• A cockpit you can rearrange. TRANSPORT, PROGRAM 16:9, VERTICAL 9:16, LEVELS, DIRECTOR, MODES and MUSIC are now separate dockable sections. Drag any of them anywhere, drop one onto another to group them as tabs, or tear one off into its own window. The arrangement is remembered between runs and ⦲ LAYOUT restores the factory placement.
• Fixed: window buttons covering the transport. SETTINGS and Setup AI were pinned to the top-right corner and overlapped the transport row as it grew. They now flow with the rest of the bar, and the status bar sits below every section where nothing can hide it.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Live AI subtitles straight to YouTube. The live translator’s lines can now be sent to YouTube as a real caption track via the official “POST captions to URL” method. Paste the signed captions URL from YouTube Studio and the app forwards each committed line as you speak. One track; viewers can auto-translate it to any language. No extra API cost, and it runs on its own thread so it never touches the render. Choose the translated line or the original transcript.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• YouTube 9:16 + 16:9 at the same time (dual stream). New YouTube 9:16 (dual) target: paste YouTube's second (vertical) stream key. The app then pushes your 9:16 canvas to it alongside the 16:9 one. One broadcast, horizontal on desktop and native vertical in the Shorts feed.
• Facebook vertical, signalled correctly. New Facebook 9:16 target for a separate portrait broadcast. The 9:16 feed now goes out as true 1080×1920 with an explicit portrait aspect, so Facebook no longer “fits to frame” / stretches it.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• No more live frames leaking before the intro. Streams used to start a couple of frames ahead of the intro graphic, briefly showing you on air before the opener. Now every stream begins on the intro's first frame, in lock-step with the recording. The opener always covers frame one.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• START is never greyed out anymore. The main transport button stays clickable at all times. While you're frozen on an outro end-card, pressing START now resumes the show via the INTRO, like the dedicated button.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• OUTRO / INTRO button moved to the main bar. The on-demand graphic trigger is now right on the top transport strip (next to START / PAUSE / NDI) instead of tucked inside the DIRECTOR tab. One click plays the OUTRO and freezes its last frame, the button flips to INTRO, and the next click resumes the show. Always at hand.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Live FPS badge on the preview. The indicator overlay can now show a colour-coded frame-rate readout (green above 50, amber above 30, red below). Add it, reorder it, corner it and scale it like every other badge. It shows on the fullscreen preview too, and stays operator-only: never recorded, never streamed.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Stability fix: no more freeze when stopping around an intro/outro. A crash-dump analysis traced an occasional hang to the intro/outro engine. Stopping while a stinger was playing could block the UI thread waiting on a decoder that a stuck ffmpeg pipe never released. The engine now force-cancels those reads, so tearing a stinger down always returns instantly. The app no longer wedges (and the safety watchdog no longer has to restart it).

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Fire your intro/outro on demand, mid-stream. A new button in the DIRECTOR panel plays the OUTRO and freezes on its last frame (an end-card, with the usual audio ducking). Meanwhile, the stream and recording keep running. Press it again to play the INTRO and resume the live show. The transport START also resumes from the frozen end-card.
• Completely independent of recording: it’s a “be right back” bumper you can trigger any time without stopping the take.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Fixed: the desktop (Chrome, cursor) could stutter while the stream stayed perfectly smooth. The app’s heavy GPU work (4K captures, mip chains, compositor passes) used to queue at the same scheduling priority as Windows’ own desktop composition. DWM missed refresh deadlines and the monitor felt choppy even though the encoded output never dropped a frame. The app’s GPU submissions now yield to the desktop, so browsing and the mouse cursor stay fluid during heavy compose. The stream’s own pacing is unaffected (verified: identical throughput and clean takes).
• Escape hatch: gpu.thread_priority in config (-7..7, default -2; 0 restores the old behaviour).

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• One START button for everything. The top transport bar is now a row of tick-boxes (REC plus each streaming platform) and a single START/STOP. It fires every ticked target together (recording + all selected streams).
• Only configured targets show up. A platform tick-box appears in the top bar only once its stream key is set in OUTPUTS. No dead buttons. Ticks stay in sync with the OUTPUTS tab, and platforms that are on air glow cyan.
• Failures speak up at the button. If a ticked target refuses to start (missing ingest server, ffmpeg/TLS trouble), the reason shows right next to START for a moment. It also stays in OUTPUTS.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• One unified chat from every platform you stream to. Paste the live links for YouTube, Twitch, TikTok and Facebook in the CHAT section and every viewer comment flows into a single merged feed. Each bubble on the canvas carries its source badge (“TW · nick”). The AI co-host, moderation and viewer memory read the whole stream as one chat.
• No keys, no OAuth. YouTube speaks the public popout-chat protocol, Twitch joins anonymously over IRC, TikTok and Facebook are read straight from the live page by the built-in browser. A “log in window” mode handles consent walls / logins once: the session is remembered.
• Live-verified. Twitch and TikTok capture was verified against real live rooms before this release. Per-platform status lines under START CHAT tell you exactly what each source is doing.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Multistream: tick the platforms, press one STREAM button. YouTube, Twitch, TikTok and Facebook Live now have selection checkboxes and a single master STREAM button. It starts every ticked platform at once (press again to stop them together). Each platform keeps its own canvas format (16:9 / 9:16), resolution, bitrate and key.
• Facebook Live connects now. The default rtmps://live-api-s.facebook.com:443/rtmp ingest is pre-filled, and the server/key join is immune to stray slashes. Pasting the full rtmp(s)://SERVER/KEY line into the key field also works.
• Streams tell you why they failed. A clear red notice under the STREAM button explains a refused start (missing key, missing ingest server, or ffmpeg/TLS trouble) instead of failing silently.
• Fixed: the app could linger in the background after closing and block the next launch (“already running”) until killed in Task Manager. A shutdown watchdog now guarantees the process fully exits even if a browser-engine or driver teardown wedges.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• YouTube Music, built into the app. Pick YouTube Music as the music source and a dedicated in-app browser window opens for login (remembered across sessions) and playlist picking. The sound goes straight into the app’s dedicated MUSIC channel: music gain, auto-ducking and beat-reactive visuals all apply, exactly like the built-in player. The music keeps playing while the window stays hidden. The player strip gets a scrub bar, previous/next controls and the on-canvas now-playing card shows the page’s cover art.
• Auto music now starts every take from the top of the track. While you are idle the opening of the current track is pre-buffered silently. So the moment you hit REC, the music begins from its first beat: gap-free, no stutter. At app launch, auto-record and auto-play now wait for the music player to finish loading before the take starts.
• Fixed: built-in folder music was briefly audible at launch when an external music source was selected.
• Fixed: a tonal feedback loop from the embedded browser. Web-player audio no longer leaks to the speakers alongside the MUSIC channel, so it can never re-enter the mix through the system capture.
• Fixed: “Leave site?” dialogs could hang the app at shutdown (and even block a Windows restart). Every page dialog from the embedded browser is now resolved automatically: nothing modal, ever.
• Smarter cuts from the AI director. The face side is decided before a mode cut (no more re-composing seconds after the switch). The camera inset lands directly in its target corner on every cut.
• Fresh LoFi + Ambient music pack. 51 royalty-free tracks in two curated folders, each with full titles and cover art. The now-playing card shows real artwork out of the box. The installer is also ~120 MB smaller.
• The short-lived Spotify source was removed: Spotify’s Web DRM (Widevine) cannot run inside embedded browsers. The built-in player and YouTube Music remain.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Hand tracking now has a real master switch, and it defaults to off. A visible HAND TRACKING toggle sits right under the mode tiles. While it is off, the hand-detection models are never even loaded (zero VRAM, zero GPU shader compilation, zero per-frame work). Every hand option, including the HANDS mode tile, is greyed out as well. Flip it on and the vision engine reloads with hand detection in about a second. The HANDS tile no longer switches the feature on behind your back, and the VISION DEBUG overlay respects the master too.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Fixed: stopping a recording could freeze the app on Intel/AMD graphics. On machines encoding with QSV/AMF (no NVIDIA), the encoder flush at stop could hold the UI thread for 10+ seconds. The window went “not responding” and looked like a crash (thanks for the crash logs). Stopping is now instant: the take finishes writing on a background thread, and a pulsing FINALIZING… badge shows in the transport bar. You can even start the next take right away.
• Smart Region extension 1.4.0: every video site, framed like YouTube. The extension now runs a tiny probe inside cross-origin player iframes (Dailymotion's own site player, Facebook / YouTube embeds anywhere). So real playback state and cursor-on-video reach the app instead of a guess. Auto-Cinematic triggers correctly on Dailymotion, Facebook, X, TikTok and Instagram.
• Tight player framing. Zooms now target the actual player card (controls included) rather than a bare video box, using a new ancestor-fitting pass. Verified live on TikTok and X.
• Smarter pick in endless feeds. When several videos are on screen, the extension prefers the one that is playing, near the viewport centre and under your cursor. This matches how feed UIs behave.
• Page-background videos no longer count. Ambient hero loops that merely decorate a page (negative z-index, aria-hidden, pointer-through, or buried under the page's own content) are now recognised as scenery. They never trigger Cinematic or a zoom. A real player's own controls and overlays still pass, verified on YouTube, TikTok and live.hexart.io itself. Grab the updated ZIP from the homepage or {app}\extension\smart-region.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• CINEMA mode dims the house lights. The camera backdrop behind the film is now properly dark and heavily blurred: presence without detail. So the playing video owns all the light (the video itself stays pixel-verbatim, untouched).
• Smart Region extension 1.3.0: X, TikTok & Instagram. The extension now sees players mounted in shadow DOM (TikTok). It keeps tracking your cursor even on sites that gag mouse events (X / Instagram) via four-way pointer listeners. It also recognises X / TikTok / Instagram embeds on other pages. So playback and cursor-on-video detection (and with them auto-Cinematic) work across all three. Grab the updated ZIP from the homepage or {app}\extension\smart-region.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Camera PiP corner moves are now choreographed, and yours to tune. A precise PiP travel speed slider (0.25–3×) sets how fast the inset journeys between corners. Three corner-change styles roll a weighted dice on every move (exactly like the montage cut-style weights). The first is classic slide across the canvas. The second is slide out (exits past its nearest edge and re-enters from the new corner's side of the frame). The third is zoom (melts to a dot on a smooth curve, grows back in the new corner). Weight 0 disables a style.
• Fully wired everywhere: tooltips on every control, Setup AI knows all four parameters, and they are MIDI / Stream Deck / hotkey-nudge targets (pip_move_speed, pip_anim_*).
• The AI commentator now has manners. It never opens its mouth while you are talking. A spoken line waits for a clean gap in your speech, and is dropped after 20 s of nonstop talking. If you interrupt it mid-sentence, its voice is cut within a fraction of a second. You always win the floor. Both behaviours are on by default and toggleable in COMMENTATOR (Setup AI knows them too).

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Settings tabs stay pinned. The tab strip at the top of the SETTINGS window no longer scrolls away with the page. It is permanently docked, and only the content of the active tab scrolls. Each tab also remembers its own scroll position now.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Separate ElevenLabs voices: subtitles narrator vs. AI commentator. The commentator now has its own complete voice setup (voice, TTS model, stability, similarity, style, speed and speaker boost). It is in a clearly labelled COMMENTATOR VOICE section, fully independent from the narrator that reads the subtitles (which keeps its own settings in SUBTITLES). Choosing "(use the subtitles voice)" still inherits everything with one click.
• Another whole-desktop freeze eliminated (file dialogs). Forensics on the 21:53 hang dump showed the system file picker's modal loop running on the render thread. That could wedge the Windows compositor on multi-GPU rigs. All file/folder dialogs now run on their own thread. The app keeps rendering and encoding while a dialog is open, and the desktop can no longer be taken down by one.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Browse buttons everywhere a file path is typed. Every "type a path" field now has a system file picker next to it. This includes the commentator's PROJECT CONTEXT .md (picking applies instantly, no extra "Set" click). Folders, images, LUTs and audio files already had theirs.
• Pick AI models from the live OpenRouter catalogue everywhere. Every model field now opens the shared model browser (search, audio/free filters, price sort). The music director and the AI commentator were the last two typing-only fields; now all six model inputs have Browse.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Privacy Mode reprogrammed (v2). The window stack is now snapshotted in the same frame as the mask math and walked once top-down. Every surface above a protected window carves through, including when the private window is the active one. A maximized private app under the control deck used to frost the whole monitor with trailing smears. Secrets stacked on secrets mask only their visible slivers, and windows hidden from capture (anti-mirror deck/projector) correctly do not punch holes. The mask geometry proves itself with 8 built-in scenarios at every launch.
• Freeze/crash while resizing the SETTINGS window: fixed for real. Forensics on the latest hang dump showed the previous async fix was a no-op for same-thread windows. Window move/size requests from the UI layer now go to a dedicated executor thread. The render thread structurally cannot get wedged against the Windows compositor anymore. This keeps the app (and your desktop) alive no matter how the window is dragged.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Free means free now. The free version has no time limits at all: every feature, unlimited hours. The output carries a small live.hexart.io watermark, and a one-time license removes it forever. The old 8-hour counter is gone from the app and the site.
• A proper EULA & safety statement. The installer now shows the full license agreement (also at live.hexart.io/eula.html, EN + PL). It covers what you may do and what data the app does and does not send. It also has a plain-language "is this safe?" summary: no malware, no ads, your recordings and API keys never leave your computer.
• Installer redesigned. New holographic wizard artwork and a welcome page that states requirements honestly (NVIDIA / Intel / AMD, not NVIDIA-only). Live progress lines tell you exactly what is being unpacked (Chromium engine, AI models, FFmpeg, music pack).
• Installer is faster. Extraction switched to fast non-solid compression: the progress bar moves from the first second and installs complete noticeably quicker. A blocking message box that used to hide behind the splash (making setup look hung) is gone.
• Groundwork for a signed installer (Azure Artifact Signing). The SmartScreen "unknown publisher" screen is on its way out once the publisher verification completes.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• GAME mode has a cursor again. The screen is captured without the OS pointer (the holographic marker replaces it). GAME mode used to switch that marker off, assuming every game draws its own. Hardware-cursor games, players and slideshows ended up with no pointer at all. The marker now stays on in GAME.
• The marker mirrors the real pointer everywhere. When an app hides the system cursor (fullscreen video, a game rendering its own sprite, mouse-look), the holographic cursor melts away. It is gone within a tenth of a second and pops back the moment the pointer returns. That way it never ghosts over nothing and never doubles a game's own crosshair.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Privacy Mode rebuilt: it frosts exactly what's visible, not the whole screen. The mask for a protected window is now carved against the complete window stack. This includes borderless, untitled and tool windows that the old logic couldn't see. A private window sitting under other windows only gets its visible slice blurred; anything covering it shows through crisply.
• Window-over-window masking fixed. Occlusion now uses live Z-order. The taskbar, always-on-top players and the projector correctly punch through, even when the private window is the active one. Previously an active private window frosted its full rectangle (a maximised one: the entire monitor).
• Exact process matching. Blacklist entries match the process name exactly (with or without .exe) instead of "contains", so guarding code no longer blurs vscode and friends.
• Privacy masking now proves its own geometry at startup: a 7-scenario self-check runs on every launch and reports into the log.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Desktop-freeze on resizing the SETTINGS window: root-caused and fixed. A hang dump showed the main thread parked inside SetWindowPos (called by the multi-viewport sync), deadlocking against the Windows compositor. This happened on multi-GPU / multi-monitor rigs with live captures. All ImGui-driven window moves/sizes now go through SWP_ASYNCWINDOWPOS (queued, never waiting on DWM). The detached windows got real OS frames, so interactive resize runs in Windows' own loop. A re-entrancy guard hardens the keep-alive tick. The freeze scenario is structurally impossible now.
• Main UI anchoring fixed for multi-viewport: the app never spawns "outside its own panel" again.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• The preview now owns the main window. The whole settings sidebar moved into its own OS window (opened with the new SETTINGS button, hidden at launch). You can drag it to any monitor. The preview stretches across the full width of the app.
• Mode panel redesigned as clickable tiles. Every director mode is a square tile with its live keyboard shortcut on the face (rebinding updates it) and a hint on hover. The tiles show activity colours: gold = on air, amber = pinned by hand, green AUTO = the autopilot is driving.
• Setup AI opens in its own window too. Chat with the agent beside the app instead of a blocking in-app dialog.
• The licence banner above the preview is gone; licence status & activation now live in a proper LICENSE tab inside the settings window.
• Cinematic got strict, window modes got calm. CINEMA now engages only while a video is playing and the cursor sits on the player. Both are detected by the Smart Region extension (v1.2.0) from inside the page. Everywhere else the browser is framed like a normal window: the extension's region never hijacks Hybrid/Working/Focus zooms any more. It only trims the tabs/address bar out of the shot while your cursor is on the page. Ride up onto the browser chrome and the full window returns.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• New "PiP punch" slider (DIRECTOR > Window reactivity to music): dial exactly how hard the camera inset micro-zooms on the beat. It ranges from perfectly still (0) through the classic amount (1×) up to double punch (2×), independently of the other windows. Also available as a MIDI / Stream Deck / hotkey-nudge target and via the Setup AI.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• A/V sync hardened end-to-end. Transient disk / ffmpeg / stream stalls used to silently drop queued video packets. Because the video timeline is packet-counted, every drop became a permanent, growing audio-video drift. The pipeline now applies real backpressure instead of dropping (encoder slows, the scheduler backfills, sync stays glued). Failed encodes return their timeline slots for re-scheduling, and the audio path got 10× more headroom. Emergency drops (destination dead for 8 s+) are loud in the log instead of silent.
• Steady frame rate on throttled setups. Windows 11 quietly ignores the app's 1 ms timer and eco-throttles it once the window is minimized (e.g. recording with the panel in the tray). The master clock then jittered in 15.6 ms steps. The app now opts out of both throttles and paces frames with a high-resolution waitable timer, precise regardless of window state.
• New drift telemetry: every recording/stream reports a growing A/V gap in the log (60 ms threshold) before you would ever hear it.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Conversation rooms fixed on every machine. The media bridge (the headless WebRTC engine that publishes your program into a room) now ships inside the installer as a self-contained runtime. Previously it silently required Node.js to be installed. So rooms only published on machines that happened to have it ("media bridge is DOWN" everywhere else). No extra installs needed anymore.
• Recording & streaming now work on Intel and AMD GPUs. Encoding no longer requires an NVIDIA card. When NVENC is unavailable, the app auto-matches a hardware encoder to your GPU (Intel Quick Sync, AMD AMF) through the bundled ffmpeg. CPU x264 is the safety net. Each candidate is probe-tested on your machine and HEVC gracefully falls back to H.264 where hardware HEVC is missing. No configuration needed (outputs.encoder can force a family).
• Pick your microphone. New microphone selector in the AUDIO tab lists every active capture device. Switching applies live: the voice chain hops to the new device in a few milliseconds, no restart. The choice persists across sessions ("default microphone" follows the Windows default communications mic, as before).
• Under the hood the live device swap (microphone and system-source alike) is now fully thread-safe against the running mixer.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Cinematic now follows your cursor. Auto-Cinematic engages while the cursor sits on a playing browser video and exits the instant the cursor leaves the player (or the video stops). It no longer lingers on the film after you have moved on to other work. A short hover dwell prevents accidental flips when the cursor merely passes across a player. Forcing CINEMA by hand still holds unconditionally.
• Cinematic is now a pure player. The whole post-processing chain (exposure adaptation, vibrance, sharpen, LUT / colour grade, vignette, scanlines, dither, bloom, chromatic aberration, VHS) is bypassed in Cinematic. Captured pixels reach the output verbatim.
• The cinematic 3D slider finally does something real: 0 = dead-flat frameless playback, 1 = a strongly tilted holographic panel with its frame, glow and particles back. The old range was so subtle it read as broken.
• New RECORDING FOLDER button in the DIRECTOR tab opens the recordings folder straight in Explorer. Each take lives in its own MLRC_<timestamp> subfolder there.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• CINEMATIC mode fixed. Forcing Cinematic by hand now stays in Cinematic. Previously it silently fell back to Working within a fraction of a second unless a perfect smart-region signal was live. The mode looked dead. The cursor-spotlight fallback now applies only to the automatic browser-video engagement, as intended.
• Auto-Cinematic works without the global smart-crop toggle. The screen-intelligence now runs whenever auto-Cinematic is armed, so a playing browser video can be seen and take the stage.
• Mixed-DPI coordinates solved for real. The Smart Region extension (v1.1.1) sends a calibration frame with every report. The app maps the region into the browser window's true physical rect, so regions are pixel-correct on monitor sets with different scale factors. This was verified end-to-end with a deliberately mis-scaled synthetic report. The extension is now also a separate download on the homepage.
• Both canvas previews (16:9 and 9:16) are now on by default on fresh installs.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• No more uncontrolled zooms in Working mode. Fixed the race where a context menu / tooltip / flyout popping up made the shot snap out to the full screen and back. The shot could also re-frame onto the popup. Ephemeral windows are now recognised and ignored. A momentary loss of the tracked window holds the current crop instead of cutting, and titled micro-popups can never steal the frame.
• Smart Regions works as intended. The browser extension's auto mode now reports only genuine media (a playing <video> or a recognised player embed). The old “largest component” fallback that zoomed into random page sections is gone (use the picker for deliberate non-media presenting). Outside Cinematic mode the smart crop engages only for real video, and every reported region is validated against the actual foreground window. So mis-projected coordinates on mixed-DPI monitor sets can never zoom the shot into an area “no window defines”. The extension (v1.1.0) now ships with the app in {app}\extension\smart-region.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• 2.5× frame rate. A deep profiling pass found and removed three hidden stalls that throttled the main render loop. The auto-exposure meter synchronously waited for the GPU every frame (~15 ms). The NVENC recorder blocked the render pipeline while the hardware finished each 4K frame (~16 ms whenever REC/stream ran). The UI/projector presents waited on the Windows compositor. On the 4K60 reference rig a full recording session went from ~23 fps to a locked 60 fps. Main-thread work went down from ~45 ms to ~3 ms per frame.
• Recording, streaming, quality and looks are bit-identical: the fixes only remove waiting, not work. New always-available MLRC_PROFILE=1 diagnostics (CPU + GPU stage timers) ship for future tuning.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• GAME MODE. When any app goes fullscreen (a game, a video player, a slide deck), the director auto-detects it and shows the app edge-to-edge. The holographic frame (borders, rounded corners, glow, 3D tilt, beat pop) dissolves away completely. Your camera rides as the standard PiP inset, with all the usual corner / shape / chroma-key / cursor-dodge rules. Auto-enters on detection, auto-leaves when the app exits fullscreen.
• New autopilot toggle (DIRECTOR > Game, on by default) and a bindable GAME action for hotkeys / Stream Deck. A smooth animated transition in and out of the mode.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Smooth screen capture restored. The 0.0.13a deadlock fix silently dropped screen frames whenever the render lock was busy. Under load the captured desktop turned visibly choppy in recordings and streams. Capture frames are now handed off lock-free and consumed at full render rate: no dropped screen frames. The Windows 10 startup-deadlock protection stays fully intact (by construction: the capture callback no longer touches any shared lock).

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• CONTROL tab, complete. Keyboard shortcuts now cover every app command (41 actions): director modes, transport, music player (play/pause, next/previous, exclude, play-order) and AI music director. They also cover subtitles & narrator, KeyCast, watermark, privacy guard, the AI master switch, theme-preset prev/next and both format previews. They are grouped into collapsible themed sections with a bound/total counter per section.
• Every slider is keyboard-steppable. New value nudge bindings: a +/- key pair for any of the 125 continuous values (5% per press, hold to repeat). They use the same registry MIDI knobs drive, so live-apply and persistence are identical.
• MIDI controllers expanded. 35 new learnable targets: subtitles & narrator (font, band, outline, TTS voice knobs) and commentator (cadence, ticker look). They also include hand-tracking stage, binaural voice, guest bus, music crossfade glide and the AI music director (weight/period/lead).
• Stream Deck. The key picker now offers the full themed action catalogue (grouped, with section headers) instead of a flat list.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Dual-format on a single GPU. The 9:16 vertical canvas now builds on every machine: one GPU or two. The 9:16 preview works and both masters record simultaneously on a single graphics card (the core one-take-two-masters promise, no dual-GPU rig required).
• Smart render management. Each format is composed only while something consumes it (its preview, the projector, REC, a live stream or NDI). An idle format costs no GPU time, so single-GPU boxes pay only for what they use.
• Status-bar render indicators. New 16:9 / 9:16 chips at the bottom show live whether each format is being rendered right now. A hint explains what wakes an idle format.
• Single-GPU 9:16 preview is now zero-copy (samples the canvas directly instead of a cross-GPU readback bridge): full rate, no added latency.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• No camera = camera truly inactive. When no camera is present (or a DeckLink feed loses signal), the camera is skipped entirely: no camera plane, no PiP inset. The AI Director never frames a camera shot. The show continues with screen-based shots only. The 0.0.15a placeholder card is gone; nothing synthetic is ever rendered.
• Signal returning (e.g. DeckLink feed back) re-enables camera shots automatically.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Recording & streaming work out of the box. FFmpeg is now bundled with the installer; no separate download or PATH setup needed. Existing custom paths keep working, and a broken path self-heals to the bundled binary.
• Clean “NO CAMERA SIGNAL” card when no camera is detected, instead of the artificial presenter avatar (removed entirely). The camera picker now honestly reports [no signal].

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Curated factory defaults. A fresh install now starts with a professionally tuned setup: broadcast audio mix, AI-director behaviour, a library of director & theme presets. No devices, API keys or machine paths are ever baked in.
• Bundled instrumental music pack (52 tracks, optional install task, on by default). It is installed into the app folder and pre-selected as the default music folder, so the music rotation plays out of the box.
• Publisher metadata. The app and installer are now signed in their version info as published by HEXART SP. Z O.O.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Critical: Windows 10 startup freeze fixed. A lock-order inversion between the Windows screen-capture present thread and the first UI paint could deadlock the app forever on launch. This was confirmed in the field on Windows 10 + GTX 10-series and proven via automatic hang-dump stack analysis. Screen capture no longer blocks the render lock.

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Hover hints on every option in every tab (180+ new tooltips). Each slider, switch and field explains what it does and what it is for.
• Setup AI knows every parameter. The assistant maps all 635 settings of the app and applies changes live, tab by tab.
• Hang auto-diagnostics. If the UI ever stalls for 8 s, the app writes a full crash-style dump and pin-points the exact stage. That way, field issues can be fixed from one report.
• KeyCast now always shows Shift while it is physically held.
• Camera list is no longer re-enumerated every frame on the SOURCES tab (lighter CPU).

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Startup diagnostics. The built-in watchdog reports the exact boot stage in the log (re-logged every 10 s during a stall), for pinpointing machine-specific hangs.
• Vision kill-switch (vision.enabled / MLRC_NO_VISION=1) for troubleshooting GPU-vision issues.
• Clear status warning when FFmpeg is missing (recording/streaming disabled until set).

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Security: the installer no longer ships any configuration file; the app writes clean factory defaults on first run. A “Reset app settings” maintenance task (on by default) heals installs poisoned by 0.0.9a.
• Room participant lifecycle. Dead sources are dropped automatically after a few seconds. Rejoins are clean (no frozen or interleaved frames), and a disconnected guest never leaves a black tile behind.
• Browser-source engine hardening for clean PCs (software rendering; no GPU-process crashes on fresh Windows installs).
• The installer tells you where the log file lives (%APPDATA%\live.hexart.io\logs).

Withdrawn: this build is no longer distributed. The current release is at the top of this page.

• Real Chromium (CEF) browser sources: exact source resolution and forced background transparency for overlay pages.
• Reset-to-default [R] chip next to every option; no more clipped buttons in the control panel.
• White-screen fix on clean Windows + GTX 10-series machines.

Download withdrawn: this build packaged a developer configuration file. Superseded by 0.0.10a.

• First public build on this site: dual-canvas 16:9 + 9:16 compositor, AI Director, holographic look and broadcast audio chain. It also includes ISO multitrack recording and streaming to YouTube / Twitch / TikTok / Facebook Live.

Download withdrawn: early build, superseded by later releases.