
davecto-cutroom
Edit your talking-head videos with Claude: a local editor, a notes loop, captions and animations in your own brand. Claude Code plugin for macOS.
Install with your AI
Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.
Install and set up davecto-cutroom (claude-plugin project) into my current project. Found on https://claudeers.com/davecto-cutroom Repo: https://github.com/davidbfrogman/davecto-cutroom Homepage/docs: — Detected install method: claude-plugin → /plugin install davecto-cutroom@davidbfrogman/davecto-cutroom Category: mcp-servers. Platforms: cli, api, web. Read the repo's README for exact setup and env vars, then install it and wire it into my project. Claudeers Health Verdict: unknown; community-verified: false. Confirm the source before running anything.
⚠ Unverified / not recently updated — review before pasting a run-this config.
/plugin marketplace add davidbfrogman/davecto-cutroom /plugin install davecto-cutroom@davidbfrogman/davecto-cutroom
git clone https://github.com/davidbfrogman/davecto-cutroom
// compatibility
| Platforms | cli, api, web |
|---|---|
| Operating systems | — |
| AI compatibility | claude |
| License | — |
| Pricing | open-source |
| Language | JavaScript |
Cutroom
Edit your talking-head videos with Claude. Cutroom transcribes a take, cuts the pauses, filler and retakes, plans a storyboard with titles, captions and transitions in your brand, and renders the finished video. You review it in an editor in your browser and leave notes on any shot; Claude makes the changes, renders a draft to check them, and replies.
Everything runs on your Mac. Your footage never leaves it.
Needs: macOS, Claude Code, and Homebrew (setup uses it to install ffmpeg, Node and uv).
Install
In Claude Code:
/plugin marketplace add davidbfrogman/davecto-cutroom
/plugin install cutroom@cutroom
Then say "set up Cutroom". Claude runs setup, which takes a few minutes the first time: it
installs ffmpeg, Node and uv with Homebrew, then Whisper (transcription), yt-dlp (YouTube) and the
headless browsers that draw graphics and animations, into ~/Library/Application Support/Cutroom.
You can also run it yourself: bash setup.sh in the plugin folder.
Use
Say "edit ~/Movies/take-3.mov with Cutroom" (or paste a YouTube link). Claude imports it and gives you the editor link.
- In the editor, select a shot and write what you want under Notes for Claude: an animation, a cut, a title, a different framing. Notes on no shot are about the whole video.
- Press Send to Claude. The header shows Claude is working; its edits appear live and highlighted, and each one is a single undo.
- Claude replies on each note: done, or a question for you. Add more notes, send again.
- Render for the final video. Exports go to
~/Movies/Cutroom/projects/<id>/exports, with.srtcaptions and YouTube chapters.
You can also edit by hand: click words to cut or restore them, drag shots, set framing, add graphics, B-roll and title cards.
Ask for animations the way you'd brief an editor: "make this line pop", "count up to 160,000 and burst", "put STARS huge behind me". Claude writes a new animation for the moment, shows you frames over your shot, and changes it when you reply with notes.
Your brand
Settings → Brand: plate, text and accent colours, a titles font and a labels font, your name and role for the lower third. Settings → Captions: caption font and size. Everything drawn into the video uses them, animations included. Make this my default starts every new video with it. Or just tell Claude: "make the captions orange on black, in my font".
Your own fonts: drop .otf, .ttf or .woff2 files into ~/Movies/Cutroom/fonts.
Your files
Everything you make lives in ~/Movies/Cutroom (set CUTROOM_HOME to move it):
| Folder | What |
|---|---|
projects/<id>/ | each video: project.json (storyboard and settings), exports/ |
fonts/ | your fonts |
music/ | music for the music picker |
broll/ | B-roll clips for every video |
luts/ | your own .cube colour looks |
motion/ | animations saved for every video |
brand.json | your default brand |
How it works
No model calls happen inside the edit. The editing decisions are rules in shared/plan.mjs;
transcription is local Whisper; everything on screen is drawn by code. Claude joins through an MCP
server (mcp.mjs): the project file is the shared state and the Cutroom server is its only writer.
Claude edits through small operations (shared/ops.mjs), every edit bumps a revision, and an editor
saving on top of a newer revision gets the new version instead of overwriting it.
The flow
- Import. Drop a file, paste a path (the file stays where it is) or paste a YouTube URL (yt-dlp). Pick the format (9:16, 16:9, 1:1, 4:5) and a Whisper model.
- Ingest (about 1× real time on an M2 for 4K, faster for 1080p): ffprobe → a 720p H.264 proxy for the browser → Faster-Whisper word timestamps and a loudness envelope → camera-cut detection → loudness measurement → a first storyboard.
- Storyboard. One row per sentence, grouped into sections. Every automatic decision is
visible and reversible:
- pauses longer than the gap setting are cut (shown as
·0.8s·in the text) - um/uh, stutters and repeated phrases are struck through (dotted = automatic)
- abandoned takes are switched off with a note saying which take was kept
- "hey guys, welcome back" openers are switched off
- list items ("First…", "Second…") become sections with chapter titles
- numbers worth seeing ("53%", "160,000 stars") get a stat callout
- "I'm Sam" gets the name lower third (your name and role, from your brand); "comment LAB" gets a CTA card
- every other same-angle cut punches in 1.1× to hide the jump; sections crossfade
- pauses longer than the gap setting are cut (shown as
- Edit. Click words to cut or restore them, ⌥-click to split a shot, drag to reorder, set framing (click the frame to set the focus point), punch-in, push/pull moves, transitions, graphics, B-roll, title cards. The preview plays the edit live from the proxy with the real LUT (WebGL) and the real graphics page on top, so there's nothing to render to check an idea.
- Render. Per-shot clips (cached, so re-renders only redo what changed) → graphics track →
assembly with transitions, music ducking and loudness → H.264 MP4, plus
.srtcaptions, YouTube chapters and the edit list.
Animations (HyperFrames)
Text animations, stinger transitions and animated title cards are HyperFrames
compositions: HTML and GSAP, rendered to transparent WebM and composited by ffmpeg. Built in:
Kinetic title, Slam text, Pop words, Invert stat (the number shows the video in
negative) and Stinger wipe. Claude writes new ones per video or for your library
(skills/cutroom/animations.md is its recipe book), timed to your words. Each takes variables
(the text, a label, a position), so one animation serves many shots, and each picks up your brand.
Behind the speaker. Any animation on a shot can sit behind you: the editor's Sits: Behind the speaker, or tell Claude. Cutroom cuts you out of the shot with Apple's Vision person segmentation (on your Mac) and lays you back over the animation, so text and shapes slide behind you, in front of the background. Titles and captions stay on top. The preview makes its cutout once per video (about 40 seconds per minute of footage on an M-series Mac); the render makes a finer one for each shot that needs it.
Looks (colour)
Built in: Paper (warm, soft blacks), Film, Clean, Ink (warm black and white).
Conversions, applied before the look: Apple Log → Rec.709 (Settings → Source colour) and
HLG/PQ → SDR, applied automatically to iPhone HDR video. Any 3D .cube file in
~/Movies/Cutroom/luts appears in the picker, with a strength slider.
What it's built from
| Job | Tool |
|---|---|
| Cuts, crops, zoom, LUTs, transitions, mixing, encoding | ffmpeg |
| Transcription with word timestamps | faster-whisper (or openai-whisper: node setup.mjs --with-whisper) |
| YouTube sources and B-roll | yt-dlp |
| Titles, lower thirds, captions | graphics/overlay.html in headless Chromium (Playwright) |
| Cutouts (the speaker in front of an animation) | Apple Vision person segmentation (swift/matte.swift, built with the Command Line Tools) |
| Animations and stinger transitions | HyperFrames 0.8.80 (lib/motion.mjs) |
| Claude Code | mcp.mjs (MCP over stdio) and lib/collab.mjs (notes, live edits, wait-for-notes) |
Files
.claude-plugin/ plugin and marketplace manifests (the MCP server is declared in plugin.json)
skills/cutroom/ what Claude reads to work with you
bin/mcp.sh starts the MCP server with Homebrew's Node
setup.sh, setup.mjs setup (Homebrew tools, then npm, Python and browsers)
server.mjs app server and JSON API (127.0.0.1 only)
cli.mjs the same pipeline from the terminal
mcp.mjs MCP server for Claude Code (no dependencies)
motion/ built-in animations (_template is the blank start)
assets/ bundled fonts and the graphics tokens (cutaways.css)
shared/ pure modules, used by Node and the browser
plan.mjs transcript → storyboard (the rules)
timeline.mjs storyboard → frame-exact timeline
brand.mjs the user's colours and fonts → CSS for graphics and animations
framing.mjs, text.mjs, ops.mjs
lib/ Node only: ingest, render, motion, cutout, collab, luts, media, deps, paths, …
swift/matte.swift the speaker matte (Apple Vision), compiled on first use
py/transcribe.py Whisper → words and loudness envelope
graphics/ the overlay page drawn into the video
ui/ the editor
tests/ vitest
Developing
npm install && npx playwright install chromium
bash setup.sh # the Python side and HyperFrames' browser
npm test
npm run dev # http://127.0.0.1:4100, restarts on code changes
node cli.mjs new ~/Movies/take-3.mov --aspect 9:16
node cli.mjs show <id> # the storyboard as a script
node cli.mjs render <id> --preview
CLAUDE.md has the rules for changing the code.
Limits worth knowing
- It edits by speech. A section with no talking (a silent demo) is cut as a pause; switch that
shot to untouched source by setting
"kind": "raw", "in": …, "out": …in project.json. - Preview transitions are approximated as crossfades or dips; the render uses the real
xfade. - 4K 10-bit HEVC decodes slowly: a 30-second 4K take renders in about 75 seconds on an M2. Quick draft (540p) is much faster for checking.
- No face tracking: framing uses one focus point per project, overridable per shot.
- Cutouts find people, not objects: a guitar or a mug in your hand may go behind the animation, and busy cushions or a second person nearby can get picked up. They need macOS 12 or newer.
// faq
What is davecto-cutroom?
Edit your talking-head videos with Claude: a local editor, a notes loop, captions and animations in your own brand. Claude Code plugin for macOS.. It is open-source on GitHub.
Is davecto-cutroom free to use?
davecto-cutroom is open-source, so it is free to use.
What category does davecto-cutroom belong to?
davecto-cutroom is listed under mcp-servers in the Claudeers registry of Claude-compatible tools.
// embed badge
[](https://claudeers.com/davecto-cutroom)
// retro hit counter
[](https://claudeers.com/davecto-cutroom)
// reviews
// guestbook
// related in MCP Servers
f.k.a. Awesome ChatGPT Prompts. Share, discover, and collect prompts from the community. Free and open source — self-host for your organization with complete…
A cross-platform desktop All-in-One assistant for Claude Code, Codex, OpenCode, OpenClaw, Gemini CLI & Hermes Agent. Only official website: ccswitch.io
🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman
An open-source AI agent that brings the power of Gemini directly into your terminal.