OpenChatCut Changelog: AI Video Editor Release Notes

This page is a readable release summary. Full bilingual history and links live in the main repository CHANGELOG.md and Releases.

[0.2.0] — 2026-08-11

Added

• Opt-in AI SDK speech routing and cloud transcription (OpenAI, Gemini, Mistral Voxtral, Cartesia, Deepgram, Groq, ElevenLabs); AssemblyAI stays the default route and on-device Whisper remains available
• In-app desktop updates: Windows and Linux builds install the next GitHub Release from the dashboard notice or Settings; macOS guides to the Releases page because the lane is ad-hoc signed
• Dashboard header shortcuts for contacting the author and opening the GitHub repository
• Opt-in blurred background fill for video and image clips: 0–100% intensity, quick presets, and edit_item backgroundFillStrength
• Visual geometry understanding: MediaPipe segmentation + face detection produce per-segment safe zones; captions auto-avoid the speaker, export QA warns on face-covering captions, auto_reframe follows the subject, and overlay graphics snap into safe zones
• edit_item source windows: sourceStartMs/sourceEndMs from media search pass through unchanged, with sourceStartSeconds/sourceEndSeconds accepted too
• Content-addressed media identity: streaming SHA-256 across every import path with deterministic relinking and deduplication, still readable by v0.1.9
• Stable caption word references with parallel source/translation lanes and deterministic CJK segmentation
• Five deterministic caption motion presets (fade-up, pop, word-pop, karaoke-pulse, none) shared by preview and export
• Server-direct external Agent editing: isolated drafts, review/commit gates, and scoped one-time upload handoffs
• Opt-in local music intelligence: hash-verified Beat This + CLAP packs for on-device BPM, beat, structure, energy, genre, mood, and instrument analysis with stale-safe beat-synced cuts
• Desktop native inference acceleration (DirectML on Windows, CoreML/Apple-silicon on macOS) with transparent fallback
• Durable Agent harness: persisted runs, approvals, checkpoints, and artifacts; reload and restart recovery; lease-fenced editing; read-only run inspector
• Opt-in SQLite project-store backend with a user-initiated, resumable migration flow
• Self-healing editor session credentials after reload

Changed

• Lower Agent token use: request-scoped tool schemas, one-shot ToolSearch expansion, bounded compaction, provider cache hints, and an in-chat usage breakdown
• Self-hosted Geist + Geist Mono UI fonts (no network font dependency)
• Semantic-index sampling now configurable per media import

Fixed

• Preserved follow-up message order in agent chats; fewer hydration round-trips
• New projects stay on the public v3 schema (v0.1.9-compatible); read-only opens no longer rewrite projects or snapshots
• Explicit cloud transcription always goes through the paid-operation approval gate; upload receipts retry until the asset edit commits; upload finalization no longer auto-starts transcription
• Watched-folder import ownership survives renderer loss; native ASR cancellation terminates the active worker immediately
• Authored clip slots preserved during relink; partial blob exports blocked; completed browser exports retained on destination reselect

View 0.2.0 release

[0.1.9] — 2026-08-06

Added

• Skills tab in the library: creative workflows + installed custom skills with search, compact cards, edit (name/summary/body), and two-step delete
• install_skill: the Agent installs a complete GitHub skill repo (SKILL.md + references/scripts/examples) into ~/.openchatcut/skills/<slug>/, with shallow-clone fallback on API rate limits
• Skills load in FULL on use (no per-file fetch), and external skills get auto dependency checks: foreign services (Codex image gen, ElevenLabs, …) map onto configured local capabilities (e.g. MiniMax image/TTS); missing ones surface with Settings guidance
• Local skill script execution (run_skill_script): whitelisted binaries run the skill's own scripts inside its directory
• Vision bypass: when the main model is not multimodal, images are described by a separate vision model; full vision model catalog + file-part input
• System proxy for server-side fetch (undici global ProxyAgent) plus an HTTPS CONNECT tunnel for the LLM proxy
• Preview source control: preview proxies are no longer auto-generated; switch between original / proxy / auto
• Official vendor icons (Xiaomi MiMo, Mureka, Fish Audio, StepFun), MCP workflow prompts, approvalMode auto sessions, and fully-automatic YOLO mode

Changed

• External MCP agents see the full internal tool surface (confirm-gated)

Fixed

• No more unfiltered-frame flash when seeking clips with WebGL effects (黑白胶片 etc.)
• Text clips not showing / half-screen video with transform keyframes
• Preview stops at the end instead of looping; humanized OOM export failures (MCP-driven, e.g. hermes)
• Browser cookies never forwarded upstream (fixes 431/400 agent-chat errors from accumulated localhost cookies)
• Clips without audio no longer fail transcription; export preflight names failing media sources

Performance

• Hardware-accelerated decoding on every video path; constant-quality proxy encoding; semantic model warm-up

View 0.1.9 release

[0.1.8] — 2026-08-06

Added

• User-visible custom skill directory (~/.openchatcut/skills/<slug>/SKILL.md) with manage_skill create, auto-discovery of hand-dropped skills, and a bundled skill-creator workflow
• Slash-command skill selection in chat (/skill:<slug> or /name), with a filtering picker, Tab/Enter activation that never overwrites the composer, and a dismissible active-workflow chip
• Agent redo (redo_last_change), named version history (manage_versions), media-pool favorite/delete/relink, auto-grade, and track reorder
• ChatGPT subscription sign-in for the built-in Agent via official Codex CLI (OAuth, model discovery, reasoning effort, dynamic tools); Claude Code subscriptions stay on local MCP
• First-class Ollama and LM Studio local Agent providers, plus StepFun / BytePlus ModelArk brains, Seedream/Seedance generation, and Inworld / Fish Audio / Speechify TTS
• Validated 4K export (2160 short edge), professional timeline tools (slip, rate-stretch, insert/overwrite, nested sequences, multicam range switch), durable generation/export jobs, scene-aware media search, and model-aware context compaction

Changed

• AI SDK 7.0.52 with longer Anthropic prompt-cache TTL and SDK-native timeouts on every LLM call site
• Creative workflows only activate (composer text is never filled); media cards are fully draggable; skills use SKILL.md + load_skill progressive disclosure
• Virtualized large library/pool/timeline surfaces, deferred Agent providers, bounded caches, and viewport-relative panel geometry
• Inspector reorganized into Basic / Video / Audio / Animation tabs; export workbench and library chrome refined

Fixed

• Large set of reliability fixes for relink-safe derivatives, ASR uniqueness, MCP session revision binding, export preflight, Chromium save pickers, atomic linked A/V overwrite/split, and preview stalls at transition boundaries

Read the 0.1.8 release story · View 0.1.8 release

[0.1.7] — 2026-07-29

Added

• Community resource packages with category previews, creator/license metadata, and install URLs shared by the website and editor
• Extension Center discovery synced with the public catalog, plus URL/file install and local enable/disable/uninstall
• Reusable resource export from the media pool for contribution
• First-run configuration guidance, direct placement onto a chosen video track, contextual clip review comments, and expanded Agent review workflows

Changed

• Streamlined resource library and Extension Center layouts; documented contribution and installation in both READMEs
• Added Ko-fi and Afdian sponsorship links in project docs

Fixed

• URL-installed packages appear immediately under Installed and remain manageable after reload
• Timeline drag keeps the playhead guide visible for captions, video, and other items

View 0.1.7 release

[0.1.6] — 2026-07-27

Added

• Precise Inspector controls with direct numeric entry, drag scrubbing, keyboard adjustment, one-click resets, and keyframe-aware editing
• undo_last_change Agent tool and per-track gap reporting in read_project

Changed

• Editing tools report compact timeline deltas instead of forcing a full project read after each action
• Frame contact sheets prefer moments where the picture changes, and editor UI styling is aligned across every major workspace
• The volatile timeline snapshot stays outside the stable Agent prompt prefix for better cache reuse

Fixed

• FCPXML media paths and transcript-edited audio segments now relink correctly in Final Cut Pro and DaVinci Resolve
• Partial Agent replies survive cancellation; reference images retain their real MIME type; media inspection failures recover consistently
• Pending autosaves flush on exit, referenced uploads survive cleanup, and invalid fades, keyframes, and clip boundaries heal or clamp safely
• Generated-result downloads retry transient failures, and semantic media search removes duplicate and weak results

Read the 0.1.6 release story · View 0.1.6 release

[0.1.5] — 2026-07-27

Added

• A colorist-grade GLSL effect suite: three-way color wheels (lift/gamma/gain), levels, highlights/shadows recovery, clarity, and an HSL qualifier for secondary correction
• Volume keyframes for audio and video clips: draw a 0–200% volume envelope with the pen tool directly on audio clips, keyframe rail on the inspector volume slider, and a volume keyframe channel in edit_item
• A change_cam agent tool for multicam switching: keep the target angle within a time range and remove the other angles' overlapping segments in one undoable batch
• inspect_color (numeric scopes: black/white points, clipping, color casts, hue histogram) and detect_beats (on-device BPM, beats, and downbeats with timeline markers) agent tools
• Named layouts (apply_layout: split screen, thirds, grid, picture-in-picture), on-device dead-air removal (remove_silence), and an in-app external MCP connection guide
• Xiaomi MiMo as a built-in agent provider

Changed

• Gemini, Kimi, Qwen, DeepSeek, and Mistral now run on their official AI SDK provider packages; Gemini speaks the native API with thought signatures handled first-class

Fixed

• Gemini no longer rejects multi-step agent tool calls with 400 "missing a thought_signature" — signatures now round-trip end to end (verified against the live API)
• Numeric enums in tool schemas that the native Gemini API rejects
• The legacy single-provider config migration no longer reroutes a newly selected provider to an old relay Base URL

View 0.1.5 release

[0.1.4] — 2026-07-26

Added

• Linux x64 AppImage desktop build in the release pipeline — the first Linux release
• Xiaomi MiMo as a built-in OpenAI-compatible agent provider

Fixed

• The collapsed thinking block now also recognizes inline <think> tags streamed by DeepSeek, MiniMax, GLM, Qwen, MiMo, and relays, uniformly across all providers
• The desktop app falls back to a random port when 5199 is taken instead of failing to launch
• Dragging a caption cue now clamps against its lane neighbors instead of overlapping them

View 0.1.4 release

[0.1.3] — 2026-07-23

Added

• Independent caption tracks, multiple caption tracks per sequence, manual captions, and video, audio, or caption type selection when creating a track
• Direct caption editing in the preview and timeline, including style drag-in, movement, and trimming both edges
• A Premiere Pro-style Rate Stretch tool that preserves the source range while changing duration and playback speed
• Model-aware Agent parameters and provider validation for image, video, music, sound, and voice generation, including expanded MiniMax and Mureka support
• OpenRouter as a built-in OpenAI-compatible Agent provider

Changed

• Moved standalone caption styling and manual editing into the Captions workspace, with a direct Caption styles entry from Transcript
• Local transcription now falls back to IndexedDB media and the original clip when extracted audio is unavailable
• Added Ctrl/Command + mouse-wheel zoom to the motion-tracking target picker

Fixed

• promptOptimizer is now sent only to MiniMax image-01
• Agent thinking renders as formatted, collapsible Markdown content
• Fixed motion-tracking previews opening on a black first frame for affected videos
• Fixed imprecise floating-point playback-speed labels and clarified how to exit Rate Stretch mode

View 0.1.3 release

[0.1.2] — 2026-07-21

Added

• WebCodecs-accelerated browser export with live progress, cancellation, and automatic server-renderer fallback
• Multi-provider stock search across Pexels, Pixabay, Unsplash, and Freesound with filters, deduplication, and partial-result handling
• Richer Agent controls for track-scoped scripts and captions, timeline targeting, exact template placement, voice isolation, and structured follow-ups
• Reusable Motion Graphic exports as ProRes 4444 MOV files with FCPXML references
• Hardware-aware local H.264 encoding, dynamic Remotion concurrency, and a heavy-export queue

Changed

• Exact template placement now retimes playback rate, fades, keyframes, zoom animation, and transitions together
• Caption sources keep a stable order; repeated Agent proposals compact only when their arguments match
• Variable-frame-rate media is normalized before Remotion playback while preserving H.264 bitrate ceilings

Fixed

• Restricted rich-widget media previews to trusted URLs
• Fixed silence markers, Motion Graphic cache collisions, and FCPXML filename mismatches
• Routed automatic export QA through a verifiable server artifact path
• Prevented concurrent exports from overcommitting local resources and cleaned up temporary files after failed or timed-out jobs

View 0.1.2 release

[0.1.1] — 2026-07-21

Added

• Configurable built-in Agent providers: Anthropic, OpenAI, Gemini, Kimi, Qwen, GLM, DeepSeek, MiniMax, Mistral, and custom OpenAI-compatible APIs
• Provider-specific API keys, base URLs, model configuration, connection checks, and model discovery
• Multi-provider runtime architecture notes and a Discord community link

Changed

• Migrated the built-in Agent runtime to the Vercel AI SDK provider abstraction
• Restricted the desktop release workflow to manual runs and reduced token permissions

View 0.1.1 release

[0.1.0] — 2026-07-20

Added

• Initial public release of the local-first, agent-native OpenChatCut video editor
• Editable multitrack projects, media management, transcript-driven editing, preview, effects, transitions, motion graphics, LUTs, and production exports
• Built-in Agent tools and MCP access for Codex and Claude Code
• Electron desktop packaging for macOS, Windows, and Linux

View 0.1.0 release

How to get updates

1. Download the latest installer from GitHub Releases
2. Source users: pull main or a release tag, reinstall dependencies, and rebuild
3. Follow discussion and previews on Discord

Format follows Keep a Changelog. Versions use Semantic Versioning.