Every build, and what changed in it.

What's new in v0.8.0

• chore(release): v0.8.0
• fix(tts): hitung durasi WAV dari header, bukan tebakan bitrate CD
• fix(packaging): kirim worker Chatterbox + addon-nya ke build terpaket
• fix(mcp): tutup celah yang membuat agen BYO diam-diam dapat suara Inggris
• feat(mcp): agen BYO bisa meminta narasi non-Inggris lewat compose_scenes
• docs: syarat hardware minimum + rekomendasi, diturunkan dari kode
• fix(tts): layer TTS baru mewarisi pilihan Voiceover dari Settings
• fix(tts): sampel preview dibacakan dalam bahasa narasi yang dipilih
• feat(tts): paket bahasa tambahan + suara Bahasa Indonesia on-device

Your project is now a folder you can carry. Export any project as a single portable bundle — composition, every asset, the raw footage it references, and the styles/presets it depends on — copy it to another PC, and import it there with one click. Plus a full internal code-quality pass: over a thousand static-analysis errors fixed to zero with the whole pipeline re-verified.

Project bundles — move a project between PCs

• Export bundle: right-click any project on the dashboard → *Export bundle…* and pick a destination. NeoMotion collects everything into one ProjectName.nmproj folder: the timeline, the project's asset bin, external raw footage referenced from anywhere on disk, and the styles, motion presets, components and doctrines the composition depends on.
• Import: the new *Import* button on the dashboard opens a bundle as a fresh project — safe to do on the same machine (nothing collides) or on a different PC (every file reference is re-linked to the new machine automatically, in every form a composition stores them).
• Dependencies install only when missing — importing never overwrites a style or preset you already have under the same name.
• Not included by design: auto-edit session state (chapters/cues), AE template packs, and the built-in sound library (it provisions itself on every install). Mac↔Mac today; Windows-side export lands later.

Under the hood

• Static-analysis errors: 1,016 → 0 across the whole codebase, with zero behavior changes — verified by the full test suite (5,195 tests), a type-check against baseline, and a live end-to-end agent QC run (project → scenes through every verification gate → assembled video).
• Fixed a class of editor-panel code that violated React's hook rules (audio, text, image, shape panels) — restructured so panels behave identically but can no longer hit render-order bugs.
• Bundle imports are transactional: a failure mid-import can no longer leave a half-created project behind.

Transcription now runs on your own machine, for free. English speech is transcribed on-device by default — no API key, no per-minute cost, nothing sent to a provider. Also: every voice picker now lets you hear the voice before you choose it, scenes stop leaving silence at the end, and your AI agent can finally edit the spoken cut itself.

Transcription runs on-device, free

• English transcription now happens on your Mac or PC by default, using the speech engine that ships with the app. No API key is needed and no audio leaves your machine for it. A cloud key stays the fallback for what on-device cannot do: other languages, an install that has not downloaded the speech model yet, or a failed run.
• Previously the on-device engine only ran when you had *no* API key configured at all — so anyone with a key kept paying per minute while a fully working local engine sat unused on disk.
• Settings now shows which engine will actually run, so you are never guessing.
• Transcription errors tell the truth now. A billing problem ("you have no credits remaining") used to appear as "Connection error." after a 30-second wait. The real reason now appears in under a second, and it names the on-device alternative.

Hear a voice before you pick it

• Play/stop buttons in every voice picker — the voiceover settings, the scene panels, the cloning screen. Choosing a narrator no longer means generating a whole video to find out.
• Voice cloning got an audition step: browse for your own audio file and hear the clone speak before you commit it to a project.
• Server-side and agent-driven builds now honor the voiceover choice you made in Settings — including the free local one — instead of quietly falling back to a paid provider.

Scenes fit their voiceover

• A scene now ends when its narration ends. Generated scenes used to keep a fixed length, so a short line left seconds of silence before the next scene. Lengths are measured from the audio itself, both ways — shrinking away trailing silence and growing when the narration runs long.

Your AI agent can edit the spoken cut

For people who drive NeoMotion from Claude, Codex, Cursor or ChatGPT:

• Speech editing by script. Your agent can read the timeline as an editable transcript, remove filler words and long pauses, cut or reorder what is said — and every caption, overlay, sound effect and speaker-frame stays locked to the right word automatically.
• Captions are editable: restyle them, highlight individual words, fix a misheard word (which corrects it everywhere at once), and export subtitles as SRT or VTT.
• New things it can make: AI video clips, background music, and sound effects — each lands in your asset library as a normal, reusable file.
• Safer project handling: deleting a project is now undoable, and one progress check covers every long-running job instead of a different one per task.
• Agents are also steered to the higher-quality scene-authoring path by default, and are stopped from silently spending your managed AI credits when they meant to use their own.

Fixes

• The red label that flashed at the start of every scene is gone. A stacked-caption feature meant for talking-head edits was being applied to every generated video, stamping the first few words of the narration over the opening seconds in a red box nobody asked for.
• Scene sliders that did nothing now work. In some AI-authored scenes the position and size controls in the Properties panel wrote values the scene ignored, so nothing moved. The app now refuses to accept a scene whose controls are not actually connected.

Your AI agent can drive NeoMotion again

• Fixed: every agent command came back as a connection error. Whether you drove NeoMotion from Claude Desktop, Codex or Cursor on the same machine, or from ChatGPT and claude.ai over NeoMotion Cloud Link, the agent could connect and see the full list of things it can do — and then every single one of them failed. Listing your projects, listing styles, starting a video: all of it.
• Nothing was wrong with the connection, the sign-in, or Cloud Link. The app reaches its own engine over a local port, and it was always asking for the same fixed port number — which is only ever correct when the app is run from source. The installed app picks a different port each launch, so it was talking to nothing. It now uses the port it is actually running on.
• This affected every installed version that offered agent control, on both platforms. It went unnoticed because a development setup happens to answer on the port it was asking for.

macOS no longer needs FFmpeg installed

• FFmpeg now ships with the Mac app. v0.5.4 could find an FFmpeg you had installed yourself; this one brings its own, so audio extraction, transcription, thumbnails and the auto-edit pipeline work on a clean Mac with nothing set up. Adds about 46 MB.
• Built from source specifically for this (scripts/build-ffmpeg-darwin.sh) under the same licence as the Windows one — no prebuilt macOS FFmpeg with that licence exists.
• Fixed: the "clip around this moment" preview in style decoding asked for an encoder that neither shipped FFmpeg has, so it had been failing on Windows outright. It now uses the same encoder as everything else.

Everything in v0.5.3, now on Windows too. The v0.5.3 Windows installer never finished building — the step that fetches the render browser used a tool the build machine doesn't have, and it failed after the download had already succeeded. Nothing about the app changed.

If you are on macOS

• This is identical to v0.5.3. You do not need to update, and updating changes nothing.

If you are on Windows

• This is the first build carrying the v0.5.3 fixes — see the v0.5.3 notes below for what they are.

Export really works now. v0.5.2 fixed the first thing that blocked exporting and uncovered the next one: the app still had no browser to render frames in, and looked for one in a place macOS does not allow anything to be written. Every export failed with mkdir '/.remotion'. The renderer's browser is now included in the app, so exporting works offline and starts immediately.

Fixed: exporting a video

• Export failed with an `ENOENT … mkdir '/.remotion'` error. Rendering runs inside a headless browser, and the app was letting the renderer choose where to keep it — which, for an installed app, resolved to a read-only system folder. It only ever worked in a developer checkout.
• That browser now ships inside the app (~175 MB), so the first export needs no download, works with no internet, and starts straight away. If it is ever missing, the app fetches it once into ~/.neomotion/render-browser instead of failing.
• A failed export no longer breaks the ones after it. Previously the first failure poisoned every later export in the same session with the same message, even after the cause was gone, and the only way out was restarting the app.

Fixed: FFmpeg on macOS

• The Mac app could not find FFmpeg, even if you had it installed. An app opened from Finder is given a minimal set of search paths that excludes Homebrew, so audio extraction, transcription, thumbnails and the whole auto-edit pipeline failed on every Mac. NeoMotion now looks in the places FFmpeg is actually installed, and if it genuinely isn't there, says so with the command that fixes it instead of failing silently. Exporting a video never needed FFmpeg and was not affected.
• Set NEOMOTION_FFMPEG_PATH to point at a specific build.

Smaller download

• The macOS installer no longer carries files only Windows can use, and the bundled tools are no longer packed twice. Together that removes roughly half a gigabyte from the Mac build — more than the render browser adds.

Export works — and the app is a third of the size. Exporting a video from an installed NeoMotion never actually worked: it failed the instant you pressed Export, in every version released so far. That is fixed. Separately, the app was shipping 4 GB of build leftovers that nothing used, so the download is now 0.84 GB instead of 1.36 GB and an install takes 2.3 GB instead of 6.3 GB.

Fixed: exporting a video

• Export failed immediately on every previous release. The app tried to assemble its renderer on your machine at the moment you exported, from files that were never included in the installer, so it stopped before rendering a single frame. It only ever worked in a developer checkout, which is why it went unnoticed.
• The renderer is now built ahead of time and shipped inside the app. Two consequences you will notice: exports start straight away instead of pausing to prepare, and the app no longer writes anything into itself while rendering.
• Verified by exporting a real project from the packaged app and checking the resulting frames.

Smaller download, much smaller install

• 4 GB of a build cache was being shipped to every user — a leftover from compiling the app that nothing reads at runtime. It never showed up in the download size because it compresses well, which is how it survived several releases.
• Download 1.36 GB → 0.84 GB. Installed size 6.3 GB → 2.3 GB.
• Automatic updates are still full downloads, so this is a saving you get on every future update, not just this one.

Known limitations

• Windows builds are unsigned (SmartScreen "Unknown publisher"). macOS is signed and notarized.
• Updates are full downloads, not patches.
• On macOS, ffmpeg is not bundled: features that read or cut your own footage (importing video, extracting audio for transcription, the talking-head pipeline) need ffmpeg available on your system. Exporting is unaffected.
• Local voices are English-only; other languages still use a cloud provider.
• Local ASR covers English; Indonesian transcription still uses the cloud path.
• The speaker cutout is baked from the full source frame, so it will not follow an active crop or push-in on that segment.

v0.5.0 would not start. It closed immediately on first launch, on a missing internal file. This release fixes that, repairs the app icon, and adds controls for the speaker card in talking-head edits — where it sits, what shape it is, and when it fills the stage.

On v0.5.0? Download this one by hand. The crash happened before the updater could run, so v0.5.0 cannot fetch this release by itself. Install it from the Download page as usual; from v0.5.1 onward, updates arrive on their own again.

Fixed: the app closed on launch

• v0.5.0 quit on startup looking for a file that was never packaged. The installer shipped the auto-updater but not everything the updater itself needs, so the app died before it drew a window.
• The packaging step now works out everything the app actually loads at runtime and ships all of it — including cases where two parts of the app need different versions of the same component. Several such gaps existed beyond the one that caused the crash, and none of them can reach an installer now.

Fixed: the app icon

• The NeoMotion icon was scrambled at small sizes — a block of coloured noise in Finder's list view, the Get Info header and the Windows taskbar, while larger sizes looked correct. Every size is now generated from the same source artwork, and Retina sizes that were missing entirely are included.
• macOS caches app icons aggressively. If the old broken icon lingers after installing, log out and back in — the icon in this build is correct.

Speaker PiP controls

• The speaker card in a lecture-style edit is now yours to place. A new Speaker PiP section in the video Properties panel: Position (any of the four corners, keeping its size and margins), Border Shape (rounded, square or circle) and Full Stage — leave it automatic, keep the speaker small the whole way, or keep them full-frame throughout.
• The canvas now shows the real card, matching where it will actually render, including the way it grows toward full-frame during full-stage moments. It used to draw a fixed placeholder box that did not correspond to the output.
• A badge marks footage where a face was found, on the timeline row and in the Properties panel, so you can tell at a glance which shots the framing and caption placement are reading.
• Available to your own agent over MCP as set_speaker_pip, with the same three controls; get_project_timeline reports the current speaker settings.

Known limitations

• Windows builds are unsigned (SmartScreen "Unknown publisher"). macOS is signed and notarized.
• Updates are full downloads, not patches.
• Local voices are English-only; other languages still use a cloud provider.
• Local ASR covers English; Indonesian transcription still uses the cloud path.
• The speaker cutout is baked from the full source frame, so it will not follow an active crop or push-in on that segment.

NeoMotion now updates itself. It tells you when a new version exists, downloads it in the background, and installs when you relaunch. Also: the "text behind the speaker" depth look, generated locally and free — and AI-designed scenes now reach the talking-head pipeline, which until now could only dress an A-roll edit from presets.

This is the last version you install by hand. v0.4.0 and earlier contain no updater, so they cannot fetch this release — download it from the Download page as usual. From 0.5.0 onward, updates arrive on their own.

Automatic updates

• A notice when a new version is out, with the version you are on shown in Settings. Until now the app never displayed its own version anywhere.
• Downloads in the background while you keep working, then a single Relaunch to update. If you ignore it, it installs the next time you quit either way.
• Turn automatic downloading off in Settings to be asked first.
• macOS updates are signed and notarized; Windows still shows the SmartScreen prompt on the installer you download by hand.
• Updates are a full download (~1.4 GB) for now — the app is not packed into a single archive, so there are no small patches yet.

Text behind the speaker

• Bake a speaker cutout from your own footage so a giant title sits *behind* the person but in front of the room — the depth look, entirely on your machine, no API key, no cost.
• Two moments are chosen automatically on every approved auto-edit: the opening title, and one emphasis beat mid-video (a spoken figure, or a punched short phrase). Both are best-effort — if a bake fails, nothing about your edit changes.
• Right-click any video layer → "Bake speaker cutout…" to place your own, with the start pre-filled from the playhead.
• Available to your own agent over MCP as bake_speaker_cutout, with the taste rules (1–2 moments per video, never recurring captions) written into the skill docs.
• Deliberately capped at short segments: this is for title moments and emphasis beats, not whole videos. Pick a stable shot — a segment that is being cropped or pushed into will not track.

Freeform scenes for talking-head edits

• The A-roll pipeline can now design each scene for the words being spoken instead of filling a preset. Same verification harness the topic director uses — gates, a determinism probe, blank and static frame checks, and an AI review of the rendered frames. A scene that fails keeps its preset layers, so a failure costs one scene, never the build.
• Your own agent can author them (list_aroll_scene_briefs / submit_aroll_scene), on your own subscription. Proven end to end on a 9m44s talk: 29 of 29 scenes authored, verified and placed.
• Verified scenes now survive a restart — a crash used to lose them all.
• A managed run stops early when your provider is out of quota instead of burning twenty minutes on retries that cannot succeed.

Auto-edit quality

• The enrichment plan survives a restart. Chapters, word-anchored cues and scene docs used to live only in memory: restarting destroyed minutes of planning and real AI spend, with no way back.
• Emphasis words are placed beside the speaker, in the emptier half of the shot, using the face track — instead of across their chest and hands — and are dressed with a rule and bracket tick so a lone word reads as a deliberate mark rather than debris.
• Empty caption chips are gone — a tier whose window outlived its words drew a bare coloured box.

Known limitations

• Windows builds are unsigned (SmartScreen "Unknown publisher"). macOS is signed and notarized.
• Updates are full downloads, not patches (see above).
• Local voices are English-only (both Kokoro and Chatterbox); other languages still use a cloud provider.
• Local ASR covers English; Indonesian transcription still uses the cloud path.
• The speaker cutout is baked from the full source frame, so it will not follow an active crop or push-in on that segment.

macOS builds are now signed and notarized by Apple — the installer opens normally instead of being blocked as "damaged". Plus an editor pass aimed at the three things that made a long session tiring — losing the playhead, hunting for the property you want, and a Scene Director board that claimed nothing was built — and MCP verbs that let an agent edit an existing video instead of rebuilding it.

macOS: installs without a fight

• The app is code-signed with a Developer ID certificate and notarized by Apple. No more "NeoMotion is damaged and can't be opened" on first launch, and no right-click → Open workaround.
• Signing now happens on every macOS build, and the build aborts rather than quietly shipping an unsigned app.
• Windows is still unsigned — expect the SmartScreen "Unknown publisher" prompt.

Editor — knowing where you are

• The timeline follows the playhead. The track view pages itself during playback, on transport seeks and on loop-back, landing the playhead 20% from the left. Scrolling by hand suspends following until the next play or scrub, so the view is never yanked out from under you. New jump-to-playhead button and an auto-follow toggle in the timeline header.
• The Scene Director board tells the truth. Cards show a real still of each scene instead of "No preview / Pending / No prompt" over finished work, built scenes get a Built badge with their layer count, and the toolbar counts Built separately from genuinely pending scenes.
• Unambiguous timecode — 00:43:02 (which read as 43 minutes) is now 0:43.02, minutes:seconds and frames, in both the transport bar and the timeline header.

Editor — faster tinkering

• Content first in the property panel. Headline text, the marked word, colours and photos now sit at the top; Transform, Effects, Expressions and Compositing move below them.
• Resizable layer-name column (140–480px, remembered per user) and a wider default, so layer names stop truncating behind the row chrome.
• Taller timeline by default — 300px instead of a height that showed about four rows — and the resize divider is now visible before you hover it.
• The Camera tab edits the scene under the playhead when nothing is explicitly selected, instead of a dead-end "Select a scene". Camera keyframes also use the scene's own clock, which was wrong for every scene after the first.
• Jump-chips in the Motion Library (Components, My Presets, Titles … Screen Motion) — the ~275-card catalog was one long scroll.

Editor — polish

• Assets have human names. Machine names (image_<uuid>, bare hashes, "Asset from stock") are replaced by the asset's own description in the bin, the detail row and the picker; newly acquired director assets are born with a real name.
• Fullscreen actually works — the preview's fullscreen button was a stub.
• Space toggles play/pause anywhere in the editor, without hijacking typing or focused buttons.
• Rotation and opacity fields no longer clip 100 into a misleading 10.
• Icon-only buttons (transport, expand chevrons, fullscreen) now have labels for tooltips and screen readers.

Connect your own agent (MCP)

• Edit an existing video instead of rebuilding it. get_project_timeline reads the timeline scene by scene, and edit_project_timeline applies surgical ops — change copy, numbers, colours, layout or an asset slot; set a duration; rename, move, remove, insert or replace a scene. Untouched scenes survive byte-for-byte, and the whole batch is all-or-nothing.
• `get_scene_source` returns a placed scene's actual source, so an agent can repair the code it wrote earlier rather than re-authoring from memory. New scene code still enters only through the existing gates.
• `duplicate_project` really duplicates. Copying a project previously cloned the row only — the copy still pointed at the original's assets, so editing one could affect the other. Assets are now cloned with their files, links are remapped, and the active doctrine is no longer dropped.

Restart NeoMotion after updating for your agent to see the new verbs — the MCP tool list is read once at startup.

Known limitations

• Windows builds are unsigned (SmartScreen "Unknown publisher"). macOS is signed as of this release.
• There is no in-app auto-updater yet — new versions are downloaded from the Download page.
• Local voices are English-only (both Kokoro and Chatterbox); other languages still use a cloud provider. Chatterbox needs a ~1.1 GB one-time model download and runs about 3.8× realtime on an Apple Silicon GPU — expect it to be slower on CPU.
• Local ASR covers English; Indonesian transcription still uses the cloud path.

The first big release since 0.1.0 — 111 commits. Headlines: voiceover and transcription now run locally with no API key, freeform (AI-designed scenes) is the default everywhere, a first-party Fox Documentary card library, and a real account + sign-in surface.

Versions 0.2.0, 0.3.0 and 0.3.1 did not reach both platforms — 0.3.1 shipped macOS only before a Windows packaging bug was found and fixed. 0.3.2 is the first release after 0.1.0 available for macOS and Windows alike.

Local AI — voiceover and transcription without an API key

• Kokoro is the new default voice engine. Voiceover is generated on your machine — no key, no per-character cost, nothing sent to a provider.
• Chatterbox voice cloning. Point it at a reference recording and it speaks in that voice. Runs on the GGML backend with Metal GPU acceleration (~8× faster than the CPU path), and the model now loads once per voice via a persistent worker instead of reloading every line.
• Local English transcription via whisper.cpp — keyless ASR that beat the cloud on word-error rate in our tests.
• The bare and whisper-cli binaries ship inside the app, with a Local AI section in Settings to manage them.

Voiceover providers

• ElevenLabs and Fish Audio join OpenAI behind one TTS registry — pick a provider and voice per project.
• A voice-reference picker for Chatterbox cloning.
• TTS failures now surface as visible warnings instead of failing silently, and every provider has a sane default voice.

Freeform mode is the default

• Every entry point now generates freeform scenes — the desktop dialog, MCP start_director, and headless callers alike. Pass generationMode: "preset" for the faster, cheaper preset pipeline.
• Every freeform scene is movable and sizable. Layout props are gate-enforced at submit time, and component + FX layers now report real gizmo bounds on the canvas.
• Verifiable asset binding — scenes bind a real assetId at submit, and assembly warns when a scene references something that isn't there.
• Updated motion taste doc (v9) driving the AI art-director review.

Fox Documentary library

• Seven first-party cards — headline, quote, timeline, photo plate, record page, lower third and newspaper page — plus eight generated paper textures, all own-artwork and installer-shippable.
• An aged-ink treatment so type stops looking digitally clean, printed on the corpus's own scanned paper.
• Composition-pattern contracts derived from a 164-pack corpus (12,533 leaves, 875 keyframe ramps), wired into every generation door so a scene picks up the style consistently.
• A scene that matches a shipped card reuses it instead of re-authoring one from scratch.

Motion Library distribution

• Pack channel — the library can grow without shipping a new app build. Motion Library → *Check for updates* fetches signed packs; publish with pnpm pack:build.
• Downloaded pack cards join the same AI casting lane as built-ins.

Account, licensing and onboarding

• Sign in with your NeoMotion account (device login) and a real Account surface showing who is signed in, with a way in and out.
• Trial status display, a welcome dialog on first run, and release gating for watermark/4K.

Connect your own agent (MCP)

• Remote access to Tier 2 — packaged shim fix, a .mcpb Claude Desktop extension, a cloudflared tunnel and the Cloud Link relay, so a hosted agent can drive NeoMotion.
• Agent skills ship inside the app and auto-install to ~/.claude/skills.
• New verbs: list_scene_drafts (recover a lost draft), generate_voiceover, composition-preview verbs, and an async scene-submit probe that dodges the MCP time limit.

Editor

• One right dock — Properties and Assets share a panel instead of stealing canvas space.
• An asset picker everywhere a media path used to be a text box.
• The dashboard shows real preview frames on project cards (and they work in the packaged app, not just in dev).
• Imported-template text is editable at the instance — no more double dive.
• Component inputs are keyframable through the property-lane system.

Export and rendering

• App-wide export tracker — progress survives leaving the project, and parallel exports no longer corrupt each other.
• Per-machine render performance settings.
• PNG-sequence renders actual frames; ProRes is reachable via mov.
• Remotion bundles are deleted when the run that made them finishes.

3D, FX and showcase

• Logo reveal — extrude a real logo into an editable 3D reveal, keep a multi-coloured logo's real colours, plus a gold-sand particle-form variant.
• Screen slot — texture a screenshot or recording onto a GLB.
• Camera-linked depth of field and a glass-floor reflection.
• Product-showcase device cards (browser, phone, laptop) on a shared stage.

Notable fixes

• Export no longer hangs at "preparing" — it spawns the real renderer.
• Settings' API keys are now reachable from server-side code and MCP.
• Chatterbox audio was hard-clipped (GGML emits int16-scale floats).
• Director bin visuals render real media, and cutouts stop dominating the frame.
• A scene's photo no longer reaches the export as a broken-image icon.
• "AI Generated" in the wizard routes to the Director Engine, not the legacy parser.
• Dialog content no longer lays out wider than its own panel.
• Cutouts are trimmed to their subject; cover-cropped images gained a focal-point knob.

Known limitations

• Builds are unsigned. Windows shows a SmartScreen "Unknown publisher" prompt; macOS shows a Gatekeeper prompt on first launch.
• There is no in-app auto-updater yet — new versions are downloaded from the Download page.
• Local voices are English-only (both Kokoro and Chatterbox); other languages still use a cloud provider. Chatterbox needs a ~1.1 GB one-time model download and runs about 3.8× realtime on an Apple Silicon GPU — expect it to be slower on CPU.
• Local ASR covers English; Indonesian transcription still uses the cloud path.

What's new in v0.1.0

• ci(release): publish desktop builds to R2 instead of GitLab job artifacts
• feat(audio): project-level audio track + serve bin-registered media outside the data root