mcp-server-documentary-generation

Status: MVP / Proof of Concept End-to-end pipeline is working. See Next Steps for planned upgrades from local inference to production APIs.

An autonomous documentary generation system orchestrated by Claude Code via MCP (Model Context Protocol). Given a topic, it produces a narrated video with AI-generated visuals — with minimal human intervention.

Built as a portfolio project demonstrating agentic AI orchestration, local ML inference, and multi-modal content pipelines.

What It Does

1. Fetches and summarises Wikipedia research on a topic (Greek-first, English fallback)

2. Parses a structured script into timestamped scenes

3. Generates image prompts per scene (Byzantine manuscript style)

4. Renders images locally with Stable Diffusion 1.5

5. Synthesises Greek narration with Chatterbox Multilingual TTS

6. Assembles everything into a video with Ken Burns effect via FFmpeg

Test output: a ~2-minute Greek-language documentary on the Fall of Constantinople (1453).

Related MCP server: comfy-ae-video-factory-mcp

Architecture

flowchart TD
    A([Topic]) --> B[research\nWikipedia API]
    B --> C[Script\nClaude inline]
    C --> D[build_storyboard\nparse scenes]
    D --> E[Claude fills\nimage prompts]
    E --> F[image_gen\nSD 1.5 CPU]
    E --> G[tts_batch\nChatterbox TTS]
    F --> H[assemble\nFFmpeg]
    G --> H
    H --> I([video.mp4])

    style A fill:#1a1a2e,color:#eee,stroke:#555
    style I fill:#1a1a2e,color:#eee,stroke:#555

MCP Tool Layer

Claude Code acts as the orchestrator. Each stage is exposed as an MCP tool that Claude can call autonomously:

Stack

Project Structure

mcp-server-documentary-generation/
├── server.py                  # MCP server — registers all tools
├── tools/
│   ├── project.py             # Folder layout helpers (slugify, scene_paths)
│   ├── research.py            # Wikipedia fetch → outline.txt
│   ├── storyboard.py          # Script parser → scenes.json
│   ├── tts_batch.py           # Chatterbox batch TTS
│   ├── image_gen.py           # SD 1.5 batch image generation
│   └── assemble.py            # FFmpeg video assembly
├── script/
│   └── aloси_1453.txt         # Test script (Greek, 5 scenes)
├── storyboard/
│   └── scenes.json            # Committed scene index with prompts
└── generated/                 # Gitignored — all media output lives here
    └── <title>/
        ├── scenes.json
        ├── video.mp4
        └── scene_XX/
            ├── script.txt
            ├── image.png
            └── audio.wav

Output Example

Topic: Η Άλωση της Κωνσταντινούπολης (1453) Language: Greek Scenes: 5 (HOOK → ΠΕΡΙΒΑΛΛΟΝ → ΠΤΩΣΗ → ΤΕΛΟΣ → ΕΠΙΛΟΓΟΣ) Runtime: ~2 minutes Style: Byzantine manuscript illustration, pencil sketch on aged parchment

Running It

Prerequisites

pip install -r requirements.txt
winget install ffmpeg  # Windows

Run a stage manually

# Research
py -m tools.research "Άλωση της Κωνσταντινούπολης"

# Parse script into scenes
py -m tools.storyboard script/aloси_1453.txt --title "Άλωση 1453"

# Generate TTS for all scenes
py -m tools.tts_batch generated/Άλωση_1453/scenes.json --title "Άλωση 1453"

# Generate images (20 diffusion steps)
py -m tools.image_gen generated/Άλωση_1453/scenes.json --title "Άλωση 1453" --steps 20

# Assemble final video
py -m tools.assemble generated/Άλωση_1453/scenes.json --title "Άλωση 1453"

Run via MCP (Claude Code)

Add to your Claude Code MCP config:

{
  "mcpServers": {
    "documentary": {
      "command": "py",
      "args": ["-m", "server"],
      "cwd": "/path/to/mcp-server-documentary-generation"
    }
  }
}

Then Claude Code can call research, build_storyboard, tts_batch, image_gen, and assemble as tools directly.

Checkpointing

Every tool skips files that already exist. You can interrupt and resume at any stage without re-running completed work:

• tts_batch → skips scenes with existing audio.wav

• image_gen → skips scenes with existing image.png

• assemble → skips scenes missing either asset

Next Steps

This MVP validates the end-to-end pipeline. Production upgrades planned:

Quality

• Images: Swap SD 1.5 CPU → FLUX Dev via Replicate API (10× better quality, seconds not minutes)

• TTS: Swap Chatterbox CPU → ElevenLabs or Azure Neural TTS (more natural, faster)

• Research: Add ChromaDB RAG for multi-source grounding beyond Wikipedia

Features

• Subtitles: WhisperX forced alignment → .srt burn-in

• Music: Overlay public domain tracks (Musopen)

• Upload: YouTube Data API v3 — auto title, description, chapters, thumbnail

• Thumbnail: Auto-generate from scene 1 image + title overlay

Scale

• Parameterise topic, language, and style via Claude conversation

• Support 15–25 min documentaries (currently ~2 min test)

• Fine-tune image prompts per historical period (Byzantine, Ottoman, Classical Greek)