YoutuberBox

Faceless YouTube video automation MCP server — shorts (9:16) and long-form (16:9).

The connected LLM does the creative work (script, scene beats, image prompts); YoutuberBox does the execution: TTS narration via voicebox, AI images via CodexImage, whisper-timed burned-in captions, and ffmpeg assembly with Ken Burns motion. Built-in quality guardrails push output away from AI slop.

Pipeline

create_project → set_script → generate_narration ┐
                              generate_images    ├→ generate_captions → assemble_video
                                                 ┘

Resource: guide://quality — the anti-slop guide (hooks, pacing targets, image prompt rules, mix levels).

Related MCP server: creatomate-mcp-server

Requirements

• Node.js 20+

• ffmpeg / ffprobe on PATH

• voicebox for TTS

• CodexImage for image generation

• Optional: mlx_whisper or whisper CLI for word-accurate caption timing

Install

npm install
npm run build

Register with your MCP client (e.g. Claude Code):

{
  "mcpServers": {
    "youtuberbox": {
      "command": "node",
      "args": ["/path/to/YoutuberBox/dist/index.js"]
    }
  }
}

Projects land in ~/YoutuberBox/projects/<id>/ (override with YOUTUBERBOX_HOME).

Configuration

Voice selection: pass voice to create_project/generate_narration as a Voicebox profile name or id; defaults to your first profile.

Captions: if your ffmpeg has libass, captions are burned in (styled, word-group pop for shorts). Otherwise they're muxed as a soft mov_text track and the tool tells you how to upgrade.

Design notes

Ideas borrowed from the best of the ecosystem: granular execution tools the LLM orchestrates (stephengpope/remotion-media-mcp) instead of a single god-tool (gyoridavid/short-video-maker's create-short-video); scene = {narration, visual spec} as the exchange format and dual-aspect maturity (MoneyPrinterTurbo); LLM-authored structure over hidden config (ShortGPT's edit-markup thesis); karaoke captions and voice discovery endpoints (short-video-maker); motion-on-stills + a cut every few seconds — the #1 documented "AI slop" complaint is static visuals.

The deliberate difference: no built-in script generation — the LLM you connect writes the script with full context of your channel and audience, and the server lints it against pacing/quality rules (hook length, words/sec, cliché openers, text-in-image prompts, static-scene warnings) instead of generating generic content itself.

License

MIT