Cadence
A prompt-native, mass-market, enterprise-scale video editor. You describe what you
want in plain language; an AI Director edits the video and shows the result —
always editable. The project is stored as edits-as-code: a declarative
EditDoc that any renderer can turn into pixels.
See AGENTS.md for the full build brief and rules, and
ARCHITECTURE.md for how it fits together.
Requirements
- Node 20+ (Node 22 LTS recommended; this repo is pinned/tested there).
- Optional, free, only for later slices: Docker (
docker compose up) and ffmpeg (video export).
Quick start (no cloud, no bill)
npm install
npm run typecheck # tsc across all packages
npm run verify # renders one real frame headlessly and asserts success
npm run verify writes a proof frame to .cadence/verify-frame.png.
Run the editor:
npm run dev --workspace @cadence/web # http://localhost:3000 (open /editor)
Real .mp4 export + Postgres persistence: docker compose up (ffmpeg baked in). Deploy to Vercel: see DEPLOY.md.
Layout
packages/core engine-agnostic EditDoc schema (source of truth) + RenderEngine contract + pure helpers
packages/render-node headless canvas RenderEngine (no browser/ffmpeg/Docker needed)
packages/render-ffmpeg pure EditDoc -> ffmpeg filtergraph plan + runExport (real .mp4)
packages/director Director tools + pure edit ops (highlight, tracks, trims, keyframes, transitions, demo, transcript, ...)
packages/understanding transcripts (Whisper CLI or offline stub) + pluggable TTS seam
packages/enhance pluggable faithful upscale providers (free / local / api / cli)
packages/db multi-tenant Postgres (migrations, pooled client, tenant-scoped repos)
apps/web Next.js editor + API (the app)
scripts/verify.ts the render-verify gate ; scripts/evals.ts capability evals
The rules (short version)
- Edits-as-code — the
EditDoc is the single source of truth.
- Never code library APIs from memory — verify from
.d.ts/docs first.
- Agentic loop — plan → smallest slice → typecheck + verify → fix → commit.
- Render-verify gate — if it doesn't render one frame, it isn't done.
- Scale by construction — client preview, splittable render, engine-agnostic ops.
- Money gate — never enable a paid/metered service without asking; free fallback first.
What you can do
A prompt-native + full manual editor (CapCut-style), all edits-as-code so preview,
canvas, and export stay in sync:
- AI: describe an edit in chat; highlight cuts, reframe, captions from speech,
edit-by-transcript, silence removal, auto-reframe, looks, slideshows, and an
interaction/walkthrough builder from screenshots.
- Manual timeline: multi-track layers with track headers (hide/show, lock,
mute/solo, rename, reorder, add/remove), drag clips between tracks, trim/split/
ripple, roll/slip/slide, an on-timeline keyframe editor, per-cut transitions,
audio fade handles, beat-sync, stickers/emoji + text presets, and a
playhead-anchored zoom.
- Rooms: Media · Edit · Words · Color · VFX · Audio · Deliver · Demo.
Full feature list: FEATURES.md. CapCut parity status:
docs/CAPCUT-STATUS.md.
Status
Social MVP + the manual/AI editor (CapCut-parity waves A–E) are built and verified;
Neon Postgres is wired and Vercel deploy is ready. See TASKS.md and
CHANGELOG.md for the slice-by-slice history.
The stub Director (free) is the default; the real Claude Director (Anthropic API,
metered) is behind a money gate and off by default — as are AI upscale and TTS.
Credits
- Emoji artwork: Twemoji (c) Twitter, Inc and other contributors, licensed CC-BY 4.0; vendored in
apps/web/public/emoji (see ATTRIBUTION.md there). Emoji names/keywords: emojibase-data (MIT).
- Fonts: see
apps/web/public/fonts/LICENSES.md.
