rakesh0x / OpenCardboard

### Chat-to-edit video editing. Local-first. Open source.
**Drop in raw footage. Describe the edit. Watch the real timeline update — cuts, captions,
beat sync, motion graphics — while your files never leave your machine.**
MIT License · macOS / Windows / Linux · Electron + React + ffmpeg
---
> The screenshot above is a real session: footage imported, transcribed locally, silence
> removed, captions generated from the transcript, and the video track re-cut to the beat
> grid of a music track — all driven through the tool API, all undoable with ⌘Z.
## Why
Most AI video tools generate fake footage or ship your files to a server. Cutboard is the
opposite:
- **Edits real footage. Never generates video.** Nothing enters the output that you didn't
 import. The agent is an editor, not a slot machine.
- **Local-first.** Originals stay where they are. Transcription, scene detection, and beat
 detection run on your machine (whisper.cpp, ffmpeg, local ONNX). Cloud models are
 opt-in per capability and clearly labeled.
- **The timeline is the source of truth.** Every agent edit is a validated operation in an
 append-only log — exact undo/redo, crash recovery, and MCP clients can edit the project
 even with the window closed.
- **One tool set, every interface.** The built-in chat and any MCP client (Claude Code,
 Claude Desktop, Codex, Cursor) drive the same 30+ typed tools.
## What it does
| | |
|---|---|
| **Footage understanding** | Word-level transcript (whisper.cpp), scene detection with keyframes, optional VLM scene descriptions — all searchable (FTS5 + local embeddings). |
| **Rough cut** | Removes pauses and repeated takes, keeps your order, reports the duration. One atomic undo. |
| **Transcript editing** | "Cut to the brushes whenever I mention them" — edit by what was said, timed in milliseconds. |
| **Captions** | Serif / bold / karaoke presets generated from word timings; styled, placed, batched. |
| **Beat sync** | Local DSP beat detection (BPM, grid, downbeats, sections) → cuts land on the grid with per-section density. Verified at ±1 frame. |
| **Motion graphics** | The agent writes sandboxed scene-tree compositions (lower thirds, counters, title cards) timed to speech. Edited only via chat, with a look-and-fix loop. |
| **Reference style** | Analyze a reference video's cut rhythm and color stats; copy the properties you choose. The reference footage never enters your edit. |
| **Effects & keyframes** | Brightness/contrast/saturation/blur with keyframed animation; speed ramps and time remap. |
| **Export** | Presets for TikTok/Reels/Shorts, YouTube, Square — hardware-encoded (VideoToolbox/NVENC/QSV/AMF) with a mixed, ducked audio track. Frame-accurate: output matches the preview. Plus OTIO hand-off to Resolve/Premiere/FCP. |
## Quick start
 [code] 
1. **Create a project** and import media (⌘I).
2. **Wait for analysis** — proxies, transcript (`✓ analyzed` badge). The first ASR model
 download is ~150 MB; manage models in **Settings → AI**.
3. **