GhostReel v0.13.2

Windows: GhostReel_*-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows offer CUDA (NVIDIA driver 570+, CUDA runtime bundled) and Vulkan (AMD, Intel or NVIDIA with a Vulkan 1.2-capable driver). Vulkan downloads have _vulkan in their filenames. Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Improved

• Indexing talks to one persistent Whisper worker instead of starting ghostreel-asr for
  every video, so the model loads once per run rather than on each clip. When transcription
  fails, the worker is released and the schedule falls back to the isolated helper it used
  before, without leaking GPU memory.
• Off-mic level measurement reads the audio the transcription decode already produced
  (mono tracks), instead of seeking the source with a separate ffmpeg per transcript line.
  On a 96-video project that stage went from 17 s to 0.6 s with identical off-mic flags.
• Speech levels now measure each stereo channel's energy before mixing, so
  opposite-phase channels can no longer cancel into a false "off-mic" reading.

GhostReel v0.13.1

Windows: GhostReel_*-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows offer CUDA (NVIDIA driver 570+, CUDA runtime bundled) and Vulkan (AMD, Intel or NVIDIA with a Vulkan 1.2-capable driver). Vulkan downloads have _vulkan in their filenames. Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• MCP tools for agents that write the edit themselves, without GhostReel's chat model:
  list_videos gives the library at a glance (category, camera, speech, what each video
  shows), get_frames returns the keyframe pictures, and get_steadiness says whether a
  stretch is shaky and where the nearest steady one is.
• get_transcript reads several videos at once and marks off-mic (interviewer) lines, or
  drops them. save_script takes dry_run to check a script without saving it, and a beat's
  bed to lay an interview's audio under muted b-roll.
• The guide explains how to add the MCP server to Claude Code, Claude Desktop, Codex,
  Antigravity and OpenCode, and lists every tool.

Changed

• search over MCP says when it fell back to keywords because the embedding model is
  unavailable.
• The website deploys after each release, from the release tag, instead of on every push,
  so the guide describes the app people can download. Its headings can be linked, and the
  "On this page" entries are links.

Added

• Vulkan builds for Windows and Linux alongside CUDA, supporting compatible AMD, Intel
  and NVIDIA GPUs with a Vulkan 1.2-capable driver. Both speech and vision/chat helpers
  use the selected backend; Vulkan installers have _vulkan in their filenames.
• GPU backend selectors on the website's download cards, with direct links and file sizes
  for each available variant. Install and source-build notes describe the Vulkan requirements.

Improved

• Qwen3.5 9B is the default for frame descriptions and script chat on all platforms,
  using GGUF with a projector on native backends and the pinned MLX bundle on Apple Silicon.
  CI validates both Qwen variants.
• CI caches compiled helpers separately by platform/backend and reuses them across release
  version bumps. Keys track helper sources, vendored bindings, transitive locked dependencies,
  toolchains, build settings and SDK versions. Vulkan SDK downloads are also cached, and
  exact helper cache hits skip native compilation and SDK setup.
• Signed updates preserve the installed CUDA or Vulkan backend using separate updater targets.
• Vulkan helper builds use portable CPU fallback settings, and staging removes stale CUDA
  runtimes before packaging a Vulkan build.

Fixed

• Windows release upload failures now trigger the existing retry handling.

GhostReel v0.13.0

Windows: GhostReel_*-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows offer CUDA (NVIDIA driver 570+, CUDA runtime bundled) and Vulkan (AMD, Intel or NVIDIA with a Vulkan 1.2-capable driver). Vulkan downloads have _vulkan in their filenames. Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• Vulkan builds for Windows and Linux alongside CUDA, supporting compatible AMD, Intel
  and NVIDIA GPUs with a Vulkan 1.2-capable driver. Both speech and vision/chat helpers
  use the selected backend; Vulkan installers have _vulkan in their filenames.
• GPU backend selectors on the website's download cards, with direct links and file sizes
  for each available variant. Install and source-build notes describe the Vulkan requirements.

Improved

• CI caches compiled helpers separately by platform/backend and reuses them across release
  version bumps. Keys track helper sources, vendored bindings, transitive locked dependencies,
  toolchains, build settings and SDK versions. Vulkan SDK downloads are also cached, and
  exact helper cache hits skip native compilation and SDK setup.
• Signed updates preserve the installed CUDA or Vulkan backend using separate updater targets.
• Vulkan helper builds use portable CPU fallback settings, and staging removes stale CUDA
  runtimes before packaging a Vulkan build.

Fixed

• Windows release upload failures now trigger the existing retry handling.

Model default

Qwen3.5 9B is now the default for frame descriptions and script chat on all platforms. CI checks both native GGUF/projector and Apple Silicon MLX variants.

GhostReel v0.12.0

Windows: GhostReel_*_x64-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows GPU acceleration needs an NVIDIA driver 570+ (CUDA 12.8 runtime is bundled). Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• Storage page (and ghostreel storage): how much space proxies, keyframes, script previews,
  chat images, exports, models and the database take, overall and per project. Each kind can be
  cleared everywhere or for one project, "Clear all caches" removes what is rebuilt on demand,
  "Clean up unused" removes files nothing uses (footage no project watches, deleted scripts and
  chats, playback proxies in the pre-0.9 format), the database can be compacted, and a project
  can be deleted together with everything made for it.

Fixed

• Deleting a project with its indexed data now also removes its playback proxies and chat
  images, and reports the space its keyframes freed.

GhostReel v0.11.0

Windows: GhostReel_*_x64-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows GPU acceleration needs an NVIDIA driver 570+ (CUDA 12.8 runtime is bundled). Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• Project search can be restricted to selected videos or a watched folder. The library supports
  checkboxes, Shift-click ranges, selecting all visible rows, and mouse drag selection in
  selection mode, with a visible search scope and a clear-selection action.

Fixed

• macOS CI creates signed DMGs directly from a staging folder, avoiding temporary
  mounted images that could fail to detach with a Resource busy error.
• Scoped keyword and semantic retrieval apply file/folder filters before candidate limits;
  unrelated footage cannot crowd selected clips out of search results. Scoped results use the
  matching folder's file path, including videos with copies in multiple watched folders.

GhostReel v0.10.0

Windows: GhostReel_*_x64-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows GPU acceleration needs an NVIDIA driver 570+ (CUDA 12.8 runtime is bundled). Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• Script-chat composer shows the tool-call budget beside the brain selectors, so the research
  limit settable in Settings is adjustable directly in chat.

• Every new script-chat answer records its actual model, token count, generation speed, and
  total time. Server rates include request time when separate decode timing is unavailable.

• AgentWorld 35B-A3B GGUF model discovery/download and a RAM-expert slider for hybrid inference.

Improved

• Local script agents allow 30 research tool calls by default, count planning replies separately,
  and inspect proposed shot and quote ranges before evidence-based draft review. Incomplete
  coverage is reported explicitly, and Settings exposes the research limit.

• Hybrid MoE inference uses a resident CPU thread pool and resident host expert buffers.
  On RTX 4070, AgentWorld Q4_K_XL measured 55.3 decode tokens/s at an allocated 128k context
  with 32 of 40 expert blocks in RAM (short-prompt benchmark, not a full-context throughput claim).

• Project workspaces use the available window width; the library gives filenames room and scrolls
  horizontally in small windows instead of squeezing metadata columns.

• Local llama.cpp inference fits GPU layers and MoE experts to available memory, supports explicit
  expert-block offload, and exposes flash attention in script Settings.

• Script chat budgets against the actual loaded context, reserves output with tokenizer counts,
  and compacts old exchanges. Text inference reuses prompt prefixes across tool rounds, including
  Qwen3.5 recurrent state via bounded checkpoints.

• Local contexts respect the model's trained limit, retry smaller allocations, and fall back to
  f16 KV storage when automatic quantized attention cannot initialize.

• Opt-in NVIDIA hardware checks cover flash attention, 32k contexts, automatic and explicit MoE
  placement, real GPU memory use, and structured script generation; validated on an RTX 4070.

Fixed

• Local agent tool calls enforce the shared argument schemas, including numeric video IDs,
  instead of accepting arbitrary objects.
• Preserve GPU projector and KV-cache offload when llama.cpp auto-fit returns its full-offload
  sentinel. Explicit MoE placement also matches fused gate/up expert tensors.

GhostReel v0.9.0

Windows: GhostReel_*_x64-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows GPU acceleration needs an NVIDIA driver 570+ (CUDA 12.8 runtime is bundled). Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• Global material search in a separate window, with project links and in-app clip previews.
• Automatic video categories, manual overrides, category filtering, and a reviewed on-disk
  organization plan that moves companion sidecars with footage.

Fixed

• Search thumbnails, keyframe strips, and script images use the local HTTP media server to
  avoid WebKit blocking asset URLs. Search falls back to the nearest available keyframe.
• Playback copies show real encoding progress, share concurrent requests safely, and reject
  cached copies whose media data fails decoding. Proxy cache keys include the configured
  vision-sampling resolution; playback copies preserve aspect ratio and use that longest edge.
• Organization refuses destination overwrites and category paths outside watched folders,
  rolls back video moves after database errors, and keeps long filesystem work off the UI thread.
• Cross-project preview navigation waits for project data, older search responses cannot replace
  newer results, and automatic category ties are deterministic.

GhostReel v0.8.1

Windows: GhostReel_*_x64-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows GPU acceleration needs an NVIDIA driver 570+ (CUDA 12.8 runtime is bundled). Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• “Check for Updates…” in the macOS GhostReel menu (⌘U), opening the existing updater
  and checking immediately.

Fixed

• Update checks and installation progress remain available when navigating between pages.
  Repeated menu clicks cannot start overlapping checks or interrupt an installation.

GhostReel v0.8.0

Windows: GhostReel_*_x64-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel). CLI bundles: available in the build artifacts.

Linux/Windows GPU acceleration needs an NVIDIA driver 570+ (CUDA 12.8 runtime is bundled). Apple Silicon uses MLX for vision/chat and Metal Whisper (macOS 14+); Intel uses llama.cpp/Whisper (macOS 12+). On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• MLX vision and script chat on Apple Silicon, with a bundled Python/Metal runtime and
  automatically selected, revision-pinned MLX model downloads. Complete model folders support
  resume, installation checks, and removal. Whisper and embeddings retain their existing engines.
• Apple Silicon model catalog for Qwen2.5-VL 3B/7B, Gemma 3 4B, and Qwen3.5 9B; the old
  Bonsai vision default migrates to Qwen2.5-VL 3B. Intel Macs, Linux, and Windows retain GGUF.

Changed

• The Apple Silicon installer now requires macOS 14+ for MLX; Intel still supports macOS 12+.

Fixed

• Multi-word searches rank exact phrases first, then matches containing every meaningful
  query term, before partial matches. Repeated partial matches cannot overwhelm a phrase when
  grouped into moments. Project and media-kind filters apply before keyword candidate limits.
• macOS signing cleanup cannot wait indefinitely on the trust service: commands have individual
  timeouts, sudo runs without prompting, and the cleanup step has a two-minute limit.

Installation notes

• Apple Silicon installers require macOS 14 or later; Intel installers require macOS 12
  or later. The MLX runtime is bundled, so no Python installation is needed.
• macOS installers remain self-signed. Drag GhostReel into Applications and approve the first
  launch in System Settings → Privacy & Security → Open Anyway.
• MLX initially describes frames serially to bound memory usage. Structured JSON replies skip
  free-form reasoning. No performance improvement over llama.cpp is claimed without benchmarks.

GhostReel v0.7.2

Windows: GhostReel_*_x64-setup.exe · Linux: .AppImage · macOS: .dmg (Apple Silicon / Intel) · CLI only: ghostreel-cli-*.

Linux/Windows GPU acceleration needs an NVIDIA driver 570+ (CUDA 12.8 runtime is bundled). macOS uses Metal. On first start open Models and download the models.

macOS builds use a self-signed certificate. Drag GhostReel to Applications; on first launch approve it in System Settings → Privacy & Security → Open Anyway.

---

Added

• macOS installers for Apple Silicon and Intel, built natively in CI. The DMG includes
  GhostReel, FFmpeg/ffprobe, and the transcription and language-model helpers; no Homebrew
  installation is needed to run the app. Requires macOS 12 or later.
• Metal acceleration for local Whisper transcription and llama.cpp vision and chat on
  supported Macs. Local helper builds select Metal automatically on macOS.
• Persistent self-signed macOS code signing, following Oriel's approach: release builds
  reuse the same certificate, sign the app and installer, and verify signatures before uploading.
• macOS automatic updates with separate signed updater archives and manifest entries for
  Apple Silicon and Intel.

Fixed

• macOS bundles include standalone, checksum-verified FFmpeg builds with the H.264 encoder,
  title rendering, and subtitle filters needed for preview rendering and playback proxies.
• Mac language-model helpers avoid a dependency on Homebrew's OpenMP runtime, and CPU fallback
  builds do not inherit the CI runner's CPU instruction set.
• CI validates both Mac architectures, required sidecar features, and system-library-only
  dependencies before packaging. Temporary signing keychains are cleaned up even on failure.

Installation notes

• Open the DMG and drag GhostReel into Applications. Builds are self-signed and are not
  notarized by Apple; after the first launch is blocked, approve the app in System Settings →
  Privacy & Security → Open Anyway.
• Existing Linux and Windows installers continue to use CUDA; macOS uses Metal with CPU fallback.