Guillotine

An AI-powered, on-device non-linear video editor for Android (also a first-class large-screen / Chromebook app).

Built with Kotlin + Jetpack Compose (Material 3 Expressive) and Jetpack Media3 (ExoPlayer for playback, Transformer for real on-device editing and mp4 export).

Your video never leaves the device. All frame/audio analysis runs on-device. Cloud AIs (Gemini/OpenAI/Anthropic/…) are controllers only: they drive the editor as text over the in-app MCP server (read the timeline, set prompts, run the on-device analysis, apply edits) and never receive your clips or frames. Cloud keys are bring-your-own, stored encrypted on-device, and there's a free, no-key on-device path (vision + an optional on-device LLM brain) so the app is fully usable with zero configuration.

Guillotine began as a web prototype (Vite + React + Express). That code has been removed; the product is the native Android app under app/. The brand assets remain in assets/.

Features

• Multi-track timeline (Vegas-style layers): import video/audio/images (SAF); split, drag across tracks; group/ungroup (grouped clips drag together). Long-press a clip edge, then drag, to trim its in/out point — a split clip re-extends back into its source the same way; its linked audio trims with it. Edge + grid snapping when placing clips, overlapping into a crossfade. Pinch to zoom width and track height, scroll through tracks, tap anywhere to seek. Clips show on-device thumbnails (video/image) and waveforms (audio).
• Multi-track compositor: the preview renders one layer per video track, stacked bottom-to-top, and crossfades a track's overlapping clips. A background-removed clip on an upper track shows the lower tracks through its matte. The exporter mirrors this (per-track sequences + crossfade + matte).
• Keyframes: animate 12 properties — opacity, scale, rotation, offset X/Y, brightness, contrast, saturation, hue, sepia, volume, and pan — with per-keyframe cubic-bezier easing. The envelope is drawn on the clip (height = value). A keyframe tool drops keyframes; tap a keyframe to select + toggle ease; drag its bezier handles to shape easing; auto-ease on by default; full cubic-bezier curve editor in the inspector.
• Crop / transform tool: pinch to scale, drag to place, twist to rotate the selected clip directly on the preview (video, image, or text).
• Text / captions: text clips are transparent overlays on video tracks — edit content + font, size/place them with the crop tool.
• Transcription → captions: generate timed, grouped caption clips from speech — on-device (Vosk, BYO model) or cloud (OpenAI Whisper). Captions burn into the export.
• Animated per-syllable captions (kinetic typography): each word is split into syllables on separate tracks with scale keyframes that grow each syllable as it's spoken — a "grow as said" effect. Ask the AI for "animated captions" or "kinetic text."
• In-app AI assistant: a minimal command bar where you type what you want and an agent drives the editor through the MCP tools. Pick any capable provider (Anthropic / OpenAI-compatible / Gemini) or the on-device LLM brain (MediaPipe LLM Inference, BYO model) — all of them are controllers that only exchange text; the actual analysis runs on-device.
• On-device vision (no video upload): free ML Kit + MediaPipe — EfficientDet-Lite2 object detection, face detection, and scene classification (~1000 ImageNet categories) — turns a prompt like "cut every frame with my phone" into split/deleted clips. Frames are sampled at 3 fps and a match extends ±5 frames, so scans are cheap. Point it at a scrubbed frame ("this is my phone") to track that specific instance via image-embedding similarity. Describe current frame gives the AI a vision readout of whatever is on screen. A free Local silence detector handles audio.
• User-defined editing tools: teach the AI named editing methods — "save this as X" or "create a tool called X that does Y" — then invoke them on any clip with "do X on this clip." The AI follows the saved step-by-step instructions using the editor's built-in tools.
• Action recorder: tell the AI "record what I do," then edit a clip by hand (split, trim, delete, keyframe, filter changes…); every action is captured. Say "save that as X" to turn the recorded steps into a reusable user-defined tool, with optional written caveats for generalization (e.g. "adapt timings to clip length").
• Generative object removal: detect + mask the object on-device, then repaint the masked frames via Leonardo.ai inpainting (BYO key); the result is a run of grouped split clips (some with generated frames) the same total length as the original.
• Background operations: analysis, generative removal, and export run in a foreground service with an ongoing progress notification — keep working while the app is backgrounded. Pause/Resume (analysis + generative) and Cancel from the notification; export is cancel-only.
• Background removal (on-device, ML Kit): segment a clip's subject and composite it over the layer below — in the live preview and baked into the export.
• Image generation: free Pollinations.ai (no key) or Leonardo.ai (BYO key, model selectable).
• Real mp4 export (Media3 Transformer): cuts removed ranges, composites every video track, positions clips on the timeline, applies per-clip filters (brightness/contrast/saturation/hue/ sepia/blur/grayscale/invert) and the crop-tool transform (scale/rotate/offset), project crop/aspect, the segmentation matte and caption overlays, bakes per-clip + track volume / pan / peak-normalize / mute / opacity, and saves to the gallery.
• Settings backup & restore: export all AI settings (provider, keys, models, speech/agent model paths, cache size) to a JSON file and import them back — handy for migrating to a new device or sharing a configuration.
• Automation (MCP): while open, the app runs a small token-gated MCP server so external AI tools (or the in-app assistant) can drive the editor. An optional end-to-end-encrypted Cloudflare relay (see tools/mcp-relay) makes it reachable from anywhere without port-forwarding.
• Whole-track controls from each track header: mute, disable/hide, volume, opacity, add clip.
• Adaptive UI: phone / tablet / Chromebook layouts, keyboard shortcuts, mouse + Ctrl-scroll zoom; Material 3 Expressive, dark with a red accent. A dropdown menu with an in-app About reader (the AzNavRail footer) surfaces this README and the privacy policy. Built-in tutorial, FAQ, and icon key (the ? button) for self-contained help.

License

See LICENSE.