🎙️ Content Creation Agent

An end-to-end AI-powered pipeline that turns any topic into a fully produced, YouTube-ready podcast video — automatically.

Built with CrewAI, FastAPI, edge-tts, Flux.1, and MoviePy, this multi-agent system researches the web, writes a natural podcast dialogue, generates voice-over audio, creates AI-generated visuals, compiles a video, and uploads it to YouTube — all through a simple API or Streamlit UI.

---

📐 Architecture Overview

Topic Input
    │
    ▼
┌─────────────────────────────────────────────────────────────────┐
│                        CrewAI Pipeline                          │
│                                                                 │
│  Agent 1          Agent 2          Agent 3                      │
│  Search Query  ──► Web Search   ──► Web Scraping                │
│  Generator         Agent            Agent                       │
│                                       │                         │
│                                       ▼                         │
│                    Agent 5  ◄──── Agent 4                       │
│                    Podcast        Podcast Script                 │
│                    Reviewer  ────► + YouTube SEO                │
└─────────────────────────────────────────────────────────────────┘
    │
    ▼
Text-to-Speech  ──►  Video Generation  ──►  YouTube Upload
(edge-tts)           (Flux.1 + MoviePy)     (YouTube Data API v3)

---

🤖 Agent Pipeline

| Agent | Role | LLM | Tools | |-------|------|-----|-------| | Agent 1 | Search Query Generator | owl-alpha (OpenRouter) | — | | Agent 2 | Web Search Agent | LLaMA 3.1 8B (OpenRouter) | SerperDevTool | | Agent 3 | Web Scraping Agent | LLaMA 3.1 8B (OpenRouter) | ScrapeWebsiteTool | | Agent 4 | Podcast Script Writer + YouTube SEO | owl-alpha (OpenRouter) | — | | Agent 5 | Podcast Script Reviewer (w/ delegation) | NVIDIA Nemotron 9B (OpenRouter) | — |

Pipeline Steps

1. Step 1 — Search Query Generation (Agent 1)
   Generates targeted, multi-angle search queries from the topic (news, expert opinions, statistics, comparisons).

2. Step 2 — Web Search (Agent 2)
   Executes each query via Serper, filters by relevance score threshold, deduplicates URLs.

3. Step 3 — Web Scraping (Agent 3)
   Scrapes full article text from each URL, cleans boilerplate, flags failures transparently.

4. Step 4 — Podcast Script + YouTube Metadata (Agent 4)
   Converts scraped content into a natural two-host dialogue with tones (curious, explanatory, enthusiastic, thoughtful, humorous), plus YouTube title, description, and SEO tags.

5. Step 5 — Quality Review (Agent 5)
   Reviews the script across 6 dimensions (naturalness, engagement, content coverage, tone variety, CTA, YouTube metadata). Can delegate back to Agent 4 for revisions up to a configurable limit.

---

🎬 Post-Processing Pipeline

After the CrewAI pipeline completes:

| Step | Component | What it does | |------|-----------|--------------| | 🔊 TTS | TexttoSpeechController | Generates per-line MP3 audio using Microsoft Edge TTS with per-host voices | | 🖼️ Image Generation | VideoGenerationController | Generates one AI image per dialogue line using Flux.1-schnell (HuggingFace) | | 🎞️ Video Compilation | VideoGenerationController | Applies tone-based Ken Burns effects, speaker overlays, and compiles MP4 | | 📤 YouTube Upload | YoutubeUploadController | Uploads the video privately via YouTube Data API v3 with OAuth2 |

---

📁 Project Structure

Content_Creation_Agent/
├── streamlit.py                # Streamlit UI (frontend)
├── README.md
└── src/
    ├── main.py                 # FastAPI app + LLM loaders + lifespan
    ├── requirements.txt
    ├── .env.example
    ├── client_secrets.json     # Google OAuth2 credentials
    ├── token.pickle            # Saved OAuth2 token
    │
    ├── controllers/
    │   ├── Agent_1Controller.py        # Search query generation agent & task
    │   ├── Agent_2Controller.py        # Web search agent & task
    │   ├── Agent_3Controller.py        # Web scraping agent & task
    │   ├── Agent_4Controller.py        # Podcast script agent & task
    │   ├── Agent_5Controller.py        # Podcast reviewer agent & task
    │   ├── CrewController.py           # Assembles all agents into Crew
    │   ├── TexttoSpeechController.py   # edge-tts TTS + audio merging
    │   ├── VideoGenerationController.py# Flux.1 images + MoviePy video
    │   └── YoutubeUploadController.py  # YouTube Data API v3 upload
    │
    ├── routes/
    │   ├── crew.py     # POST /create-script
    │   ├── audio.py    # POST /generate-audio
    │   ├── video.py    # POST /generate-video
    │   └── upload.py   # POST /upload-youtube
    │
    ├── helpers/
    │   ├── config.py   # Pydantic output schemas for all agents
    │   └── schema.py   # App settings loaded from .env
    │
    └── Results/        # Auto-generated output files

---

🔧 Configuration & Environment Variables

Copy src/.env.example to src/.env and fill in all values:

# Output directory for generated files
output_dir=Results

# API Keys
OPENROUTER_API_KEY=your_openrouter_api_key
SERPER_API_KEY=your_serper_api_key
HUGGINGFACE_API_KEY=your_huggingface_api_key
LANGTRACE_API_KEY=your_langtrace_api_key   # optional

# Models (via OpenRouter)
OPENROUTER_MODEL=openrouter/owl-alpha
LLAMA_MODEL=openrouter/meta-llama/llama-3.1-8b-instruct
NVIDIA_MODEL=openrouter/nvidia/nemotron-nano-9b-v2:free

# Agent Behaviour
num_keywords=5          # Max search queries Agent 1 generates
no_keywords=5           # Same (used via Settings)
score_th=0.5            # Minimum relevance score to keep a search result
max_results_per_query=3 # Max results per search query
min_word_count=200      # Min words for a scraped article to be kept

# Quality Thresholds
approval_threshold=7.5  # Score Agent 5 needs before approving script
revision_threshold=6.0  # Score below which Agent 5 triggers revision
max_delegations=2       # Max revision loops allowed

---

🚀 Getting Started

Prerequisites

• Python 3.10+
• FFmpeg (required by MoviePy for video export)

1. Install dependencies

cd src
pip install -r requirements.txt

2. Set up environment

cp .env.example .env
# Edit .env with your API keys

3. Set up YouTube OAuth2 (for upload)

1. Create a Google Cloud project and enable the YouTube Data API v3
2. Create OAuth2 credentials (Desktop app) and download as client_secrets.json
3. Place client_secrets.json in src/
4. On first upload, a browser window will open for authorization — the token is then saved to token.pickle

4. Start the FastAPI backend

cd src
uvicorn main:app --reload

The API will be available at http://localhost:8000

5. Start the Streamlit frontend

# From the project root
streamlit run streamlit.py

---

🌐 API Reference

POST /create-script

Runs the full 5-agent CrewAI pipeline and generates a podcast script.

| Parameter | Type | Default | Description | |-----------|------|---------|-------------| | topic | str | required | The podcast topic | | websites_list | str | required | Preferred source websites (comma-separated) | | language | str | required | Language for search queries (e.g. English) | | target_word_count | int | required | Approximate article length target | | tone | str | "informative and professional" | Writing tone | | host_1_name | str | "Alex" | Name of host 1 | | host_2_name | str | "Sara" | Name of host 2 | | target_duration_minutes | int | 3 | Target podcast duration in minutes |

Response: JSON with suggested search queries from Step 1.

---

POST /generate-audio

Converts the podcast script into per-line MP3 segments and merges them.

⚠️ Requires /create-script to have been run first.

Response: final_audio.mp3 file download.

---

POST /generate-video

Generates one AI image per dialogue line, overlays speaker text, applies tone-based video effects, and compiles the final MP4.

⚠️ Requires /generate-audio to have been run first.

Response: podcast_video.mp4 file download.

---

POST /upload-youtube

Uploads the generated video to YouTube as private (safe for review before publishing).

⚠️ Requires /generate-video to have been run first.

Response: { "youtube_url": "https://www.youtube.com/watch?v=..." }

---

🎨 Video Effects by Tone

| Dialogue Tone | Visual Effect | |--------------|---------------| | curious | Slow zoom-in (discovery feeling) | | explanatory | Pan left-to-right (steady, clear) | | enthusiastic | Fast zoom-in + fade in (energy) | | thoughtful | Zoom-out (reflective, pulling back) | | humorous | Handheld camera shake (playful) |

---

📊 Output Files

All files are saved to output_dir (configured in .env):

| File | Description | |------|-------------| | step_1_suggested_search_queries.json | Agent 1 output — search queries | | step_2_search_results.json | Agent 2 output — filtered search results | | step_3_scraped_articles.json | Agent 3 output — scraped article content | | step_4_podcast_script.json | Agent 4 output — podcast script + YouTube metadata | | step_5_podcast_review.json | Agent 5 output — quality review scores | | audio_segments/ | Per-line MP3 audio files | | frame_images/ | Raw + overlaid PNG frames | | final_audio.mp3 | Merged full podcast audio | | podcast_video.mp4 | Final compiled video |

---

🧩 Key Dependencies

| Library | Purpose | |---------|---------| | crewai | Multi-agent orchestration | | crewai-tools | SerperDevTool, ScrapeWebsiteTool | | fastapi + uvicorn | REST API backend | | streamlit | Frontend UI | | edge-tts | Microsoft Neural TTS (free) | | moviepy | Video compilation | | Pillow | Image processing & overlays | | huggingface_hub | Flux.1 image generation via HuggingFace Inference | | google-api-python-client | YouTube Data API v3 | | pydantic | Structured agent output schemas |

---

⚠️ Notes

• The pipeline runs sequentially — each agent waits for the previous one to complete.
• Agent 5 can delegate back to Agent 4 for script revisions (up to max_delegations times).
• Video generation is CPU/IO-intensive and runs in a background thread to avoid blocking FastAPI.
• YouTube uploads are set to private by default — review the video before publishing.
• Set verbose=True in agents and crew to monitor agent reasoning in the terminal.

---

📄 License

MIT License — feel free to use, modify, and extend.