What is Smashcut?
Smashcut is a local-first AI video editor for motion graphics, explainers, product launches and social videos. Instead of generating pixels you can't edit, your AI agents (the Claude Code or Codex CLI you already use) build the video as code: every title, chart, transition and music cue is a small HTML/CSS/GSAP clip or a Tone.js score that renders frame-exact to MP4.
- Storyboard first — the agent writes the script (scenes, copy, timing, visuals) before it makes anything, then works scene by scene with live statuses.
- Direct, don't regenerate — "make the number bigger", "redo scene 3", "lock what I approved". Pin comments on frames; agents resolve them.
- Everything a video needs — motion graphics, music, voice-over, captions, 3D in Blender, footage tracking, editing and export.
- Yours — runs on your machine with your own agent subscription. Projects are plain folders with git history. MIT licensed.
Install
Requirements: macOS or Linux, Node.js 20+, and at least one agent CLI you are logged into — Claude Code (npm i -g @anthropic-ai/claude-code) and/or Codex (npm i -g @openai/codex).
npx smashcut-studio # opens http://127.0.0.1:4317
# or install it
npm i -g smashcut-studio && smashcut ~/Videos/my-video
smashcut doctor # checks ffmpeg, render browser, agent logins, Blender, disk, port
ffmpeg is used from your system if present (bundled fallback otherwise). Rendering uses your installed Chrome or downloads Chrome Headless Shell once. Blender (optional) enables 3D scenes.
Features
Storyboard-first AI video production
The agent writes a script, then builds scene by scene. Every scene shows its status — not started, in progress, needs review, done, approved — and which agent is on it. Select scenes to Make, Redo, Variations or Voice-over them (you add what to change), Approve or Lock them — approved and locked scenes are off-limits for agents.
Motion graphics as code
Clips are HTML/CSS/JS (GSAP, SVG, canvas) — pure functions of time, rendered frame-exact. Flexible parameters, design-system tokens, an on-frame layout gizmo, 24 transitions, 17 effects, masks and blend modes that render identically in preview and export.
Professional craft built in
Agents follow a core craft skill distilled from measured references (HyperFrames films, BBC/Netflix subtitle specs, Material / Carbon / Fluent motion tokens, NN/g): story before frames, audio as the clock, one focal motion at a time, easing by role, read time and type sizes, safe areas for 9:16, seams between scenes, sound levels — and a review loop with contact sheets.
Techniques, references and inspiration
57 built-in techniques (each with a guide the agent studies — not templates to paste) and a small curated inspiration set — the best real-world motion pieces and edits (title sequences, launch films, data stories, ads, creator formats). Every one says in a line what we take from it and shows it: a looping mini-scene of ours that re-creates just that technique, never the piece itself. Add them to a project as references, bound to scenes.
Music, voice-over and sound
- Music as code — Tone.js scores rendered offline; beats, sections and markers drive the animation.
- Voice-over as a process — speech text with numbers spelled out, stress marks and pauses; ElevenLabs takes with a pace check so the voice doesn't speed up or slow down; captions keep the display text.
- Your own voice — a teleprompter highlights the script word by word at your pace and waits while you are silent; takes are recorded from the microphone, cleaned (noise, clicks, EQ, compressor, de-esser, loudness to target — with measured before/after), aligned to the script for captions and put on the voice-over track per scene.
- Loudness normalisation, denoise, silence removal and music ducking.
3D: Three.js by default, Blender for the heavy stuff
- Three.js clips — logos and type extruded in space, product turntables, particles that form shapes, 3D charts, isometric scenes. Live in the preview (no render step), tweakable params, transparent over footage, frame-exact in export. Agents can install any npm package into the project and use it in a clip.
- Blender — physics, cloth and fluids, photoreal Cycles light, heavy geometry: Python scenes rendered headless in stages — mock (seconds) → look (EEVEE) → final — placed on the timeline and replaced in place.
Footage tools for agents
Camera tracking (graphics glued to the world), planar tracking (screen replacement with perspective), object tracking by words, auto-reframe 16:9 → 9:16, blur faces/plates, background removal for images and video, chroma key, shot detection and rhythm analysis, morphs, slow-mo.
Built for real work
Projects are git repositories (every agent task is a commit, undo/redo), several projects in browser tabs, video versions for other languages/aspect ratios/lengths, sub-agents that build scenes in parallel, agents isolated to the editor's tools by default, export to MP4 / ProRes 4444 with alpha / WebM / GIF, English UI with Russian and more.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
