VibeTube

Desktop multicam recorder (macOS / Electron). It captures your screen and your webcam + microphone in sync, organises them into projects made of several clips, and hands the folder to an AI coding agent — Claude Code or Codex — which edits the final video for you: automatic shot selection (full-cam / screen / PiP), subtitles, HyperFrames graphics and sound effects.

Then it publishes it. The same agent reads the subtitles of the finished cut and writes five title options and a description with real chapters, and Upload-Post pushes the video straight to YouTube, TikTok, Reels and the rest — you never open a browser. See Publishing to YouTube & co. below.

You record. The agent edits and publishes.

How it works

flowchart LR
    S["🖥️ Screen<br/>(no audio)"] --> R
    W["🎥 Webcam + 🎙️ mic<br/>(the only audio)"] --> R
    R["<b>VibeTube</b><br/>N synced clips"] --> P[("Project folder<br/>clips/ + sync.json")]
    P --> A{"<b>Agent</b><br/>Claude Code<br/>or Codex"}
    A --> V["<b>video-use</b> skill<br/>transcribe → pick shots → cut"]
    V --> G["HyperFrames graphics<br/>+ HeyGen SFX + subtitles"]
    G --> H["edit/final.mp4<br/>1920×1080"]
    G --> N["edit/final_9x16.mp4<br/>1080×1920"]
    H --> U["<b>Upload-Post</b><br/>YouTube, TikTok, Reels…"]
    N --> U

Every clip keeps two tracks on one timeline: screen.webm carries the picture only, and webcam.webm carries your camera and the microphone — it is the single source of audio for the whole project. sync.json stores the offset between them, so the editor can cut between shots without ever breaking the voice.

What happens to each clip before the agent sees it:

flowchart LR
    C["Raw take"] --> B["Background replaced<br/>MatAnyone2 · on-device"]
    B --> E["Voice cleaned<br/>NVIDIA Studio Voice"]
    E --> Q["Validated<br/>frames · size · audio hash"]
    Q --> D["clips/clip_NN/"]

Requirements

• Node.js 22+ and npm
• macOS (uses Electron's desktopCapturer plus the Screen Recording permission)
• ffmpeg on your PATH (Homebrew recommended)
• For AI editing and scripts: Claude Code or Codex CLI, installed and signed in. The app reuses the configuration and authentication of whichever CLI you pick.
• For the editing step: the video-use skill with ffmpeg and ELEVENLABS_API_KEY
• For publishing: an Upload-Post API key (UPLOAD_POST_API_KEY) in ~/.config/record-studio/.env
• For voice enhancement: NVIDIA Studio Voice NIM (NVIDIA_API_KEY or NGC_API_KEY in ~/.config/record-studio/.env)

Getting started

npm install
npm start

macOS permissions: the first time, grant Screen Recording, Camera and Microphone to the app (or to Electron in dev mode) under System Settings → Privacy & Security, then relaunch. If the screen preview is black, check that permission and that the selected source still exists.

Choosing Claude Code or Codex

In the top bar, Agent → Codex switches the agent used for the next edits, iterations, style analysis and scripts. The choice persists across restarts; Claude Code is the default. Tasks and terminals that are already running stay on the agent they started with.

To use Codex, install its CLI and sign in:

npm install -g @openai/codex
codex login

Automatic editing runs codex exec --json, and the built-in terminal opens Codex interactively. Both use workspace-write with network access for research and downloads; automatic tasks never ask for approvals, while the terminal does let you approve actions. The model is inherited from your Codex configuration. See Codex non-interactive mode.

Each project keeps one conversation per provider: switching to Codex starts its own thread and you can come back to the Claude one later. The interactive terminal resumes the provider's last session in that folder. The montage brief is written to AGENTS.md for Codex and CLAUDE.md for Claude, preserving any instructions already there. Both are given the path to the video-use skill bundled with this app.

Recording

Projects tab

1. Pick the root folder where all projects will live.
2. Type a name and hit ＋ Create (it becomes the active project).
3. The gallery shows every project: clip count, duration, date, and its final edit (poster plus ▶ to open it) when edit/final.mp4 exists.

Record tab

1. Pick the screen or window. Sources are grouped by type and labelled; the selected one is highlighted. Refresh reloads the list and its thumbnails.
2. Check the screen and webcam previews.
3. ● Record → 3·2·1 countdown → talk or demo.
   • ⏸ Pause / ▶ Resume within the same clip.
   • ■ Stop closes the clip and adds it to the project.
4. Repeat: every Record→Stop adds clip_02, clip_03… to the same project.

Demoing the app itself while recording

Turn on Keep the interface visible while recording, next to the Record button. The setting persists: the main window stays open and capturable, and the small floating controls appear as usual. Turned off, the main window hides during the take. Select the screen, or the VibeTube window itself, to include the interface in the video.

You can leave the project and open Projects, Scripts or another project without interrupting the take. The recorder keeps writing to the project where the session started, no matter which screen you are showing. Its name is displayed in the floating controls.

• ⏸ / ▶ pause and resume the take.
• ■ saves the clip and leaves the small recorder ready for another.
• ● starts another clip in the original project, even while viewing a different one.
• ↺ discards the current take and re-records into that same project.
• ✓ ends the recorder session.

Camera and microphone stay available between clips while that session is open. Navigating never changes where the clip is saved, not even when in-memory buffering is used instead of streaming to disk. The session's project cannot be deleted until you end it. Duplicate starts are blocked during the countdown and while saving.

Teleprompter

The teleprompter opens in a separate window that stays out of the capture. You can load a script, edit it and restart the read. Line breaks from the original script are preserved.

Project layout

<root>/<project>/
├── project.json                 ← metadata + clip list
├── clips/
│   ├── clip_01/ screen.webm  webcam.webm  webcam_orig.webm  webcam_enhanced.wav  sync.json
│   ├── clip_02/ ...
│   └── ...
└── edit/                        ← written by video-use
    ├── final.mp4  final_9x16.mp4  final.srt  _poster.jpg
    └── publish.json              ← titles, description with chapters, tags

Handoff to video-use

Open the project folder with the video-use skill and ask for what you want:

"Edit this multicam recording: alternate between my camera and the screen depending on what I'm saying, PiP while I explain over the demo, a HyperFrames intro, and subtitles."

Continuous audio always comes from webcam.webm; only the video shot changes. The full contract (mapping several clips, multicam EDL format) is in HANDOFF.md.

AI voice enhancement (NVIDIA Studio Voice NIM)

VibeTube integrates NVIDIA Studio Voice NIM (48k-hq) over gRPC to turn microphone audio into studio-grade voice: it suppresses room echo, background noise, keyboards and air conditioning, improving presence and clarity at 48 kHz.

Ways to use it

1. Automatic after recording (background). Once a take is stopped and validated, the app processes the audio on its own. A backup is kept (webcam_orig.webm) and the clean audio is remuxed into webcam.webm without re-encoding the video.
2. On demand from the UI.
   • Per clip: the ✨ button on each clip card processes or re-processes it.
   • Whole project: Enhance audio (AI) in the clip list header.
   • Montage bar: the Studio voice (NVIDIA) switch (on by default).
3. Manually from the CLI:

   # One clip:
   python video-use/helpers/enhance_voice.py --clip clips/clip_01
   
   # Every clip in the current project:
   python video-use/helpers/enhance_voice.py --all
   
   # A standalone audio or video file:
   python video-use/helpers/enhance_voice.py --input <file> --output <enhanced.wav>
   

   The helper automatically splits takes longer than 4.5 minutes with a soft crossfade, respecting the API limits.
4. Whisper transcription (helpers/transcribe_whisper.py) prefers webcam_enhanced.wav when it exists, which gives better lexical accuracy and tighter timestamps for subtitles and SFX.
5. API key: store it in ~/.config/record-studio/.env:

   NVIDIA_API_KEY=nvapi-...
   

   With no key, or offline, the app simply skips enhancement without interrupting recording or editing.

Publishing to YouTube & co.

Once the final cut exists, the Result step can send it straight out through [Upload-Post](https://upl