Digestr

Digestr is an automated daily digest application designed to ingest content from various sources, primarily YouTube videos, summarize the key points, and generate a high-fidelity audio digest for you to listen to on the go.

The project uses Gemini for summarization and supports multiple text-to-speech backends for audio generation: Gemini TTS and Kokoro via kokoro-fastapi.

Features

• YouTube Integration: Automatically fetches captions/transcripts from YouTube URLs.
• Intelligent Summarization: Uses Gemini models to create concise, insightful summaries of long-form content.
• Daily Digest Synthesis: Weaves multiple sources into a cohesive news-style podcast script.
• Pluggable TTS Providers: Generate audio with either Gemini TTS or Kokoro.
• Multi-Interface Support: Native support for CLI, a FastAPI Backend, and an MCP server for AI agents.
• Article Ingestion (Planned): Extract text from daily news articles.

Workflow Diagram

The following diagram illustrates the end-to-end data flow of the Digestr application:

graph TD
    %% Input Sources
    In_YT[/YouTube URLs/] --> Ext_Vid[Extract Video IDs]
    Ext_Vid --> Fetch_YT[Fetch Transcripts]
    
    %% AI Processing
    Fetch_YT --> AI_Summ[Gemini Model: Summarization]
    AI_Summ --> Output_Text[News Script / Summary]
    
    Output_Text --> AI_TTS[TTS Provider: Gemini or Kokoro]
    AI_TTS --> Output_Audio[Audio Digest .mp3]

Loading

Setup & Installation

This project uses uv for lightning-fast dependency management.

1. Clone the repository and install dependencies: Ensure you have uv installed, then run:

   uv sync

2. Configure Environment Variables: Copy the example environment file:

   cp .env.example .env

   Edit .env and configure:

   • GOOGLE_API_KEY for Gemini summarization and Gemini TTS
   • KOKORO_API_URL if you want to use Kokoro TTS (defaults to http://localhost:8880/v1/audio/speech)

3. Optional: Run Kokoro locally:

   docker run -d -p 8880:8880 \
     --name kokoro-fastapi \
     ghcr.io/remsky/kokoro-fastapi-cpu:latest

Usage

You can run the CLI directly from your terminal, passing one or more target URLs via --urls. Audio generation also supports --tts-provider and an optional --voice override.

Provider defaults:

• gemini uses Puck
• kokoro uses af_bella

Single Video Summary

uv run python -m src.interfaces.cli.main \
  --urls "https://www.youtube.com/watch?v=_Hsdazxi9SI" \
  --tts-provider gemini \
  --voice Charon

Multi-Video News Program (Daily Digest)

Pass multiple URLs separated by spaces. The AI will weave them together into a cohesive podcast episode!

uv run python -m src.interfaces.cli.main \
  --urls "dQw4w9WgXcQ" "BffWWGOgcWs"

Generate Audio With Kokoro

uv run python -m src.interfaces.cli.main \
  --urls "dQw4w9WgXcQ" \
  --tts-provider kokoro

API Usage

You can start the backend server using:

uv run uvicorn src.interfaces.api.main:app --reload

Endpoints

• POST /digest/text: Returns a JSON object with the generated script.
• POST /digest/audio: Returns a streaming MP3 file of the digest.

tts_provider also affects text output. Gemini-generated scripts include expressive bracketed tags for Gemini TTS, while Kokoro-generated scripts avoid those tags and rely on punctuation.

Request Body (JSON):

{
  "urls": ["https://youtube.com/watch?v=..."],
  "tts_provider": "gemini",
  "voice": "Puck"
}

Kokoro Example:

{
  "urls": ["https://youtube.com/watch?v=..."],
  "tts_provider": "kokoro",
  "voice": "af_bella"
}

MCP Interface

Digestr includes an MCP (Model Context Protocol) server built with FastMCP, allowing AI agents to natively access its capabilities.

To start the MCP server manually or test it:

uv run python -m src.interfaces.mcp.main

You can also use the FastMCP inspector:

uv run fastmcp dev src/interfaces/mcp/server.py:mcp

Connecting an Agent

Add the following to your agent's MCP configuration (e.g., mcp_config.json or claude_desktop_config.json):

{
  "mcpServers": {
    "digestr": {
      "command": "uv",
      "args": [
        "run",
        "python",
        "-m",
        "src.interfaces.mcp.main"
      ],
      "cwd": "/absolute/path/to/digestr",
      "env": {
        "GEMINI_API_KEY": "your-api-key",
        "KOKORO_API_URL": "http://localhost:8880/v1/audio/speech"
      }
    }
  }
}

Exposed Tools

• get_transcript(url): Fetches a raw YouTube transcript.
• summarize_video(url, provider): Generates a news summary for a single video.
• create_news_program_script(urls, provider): Synthesizes a news broadcast script from multiple videos.
• generate_audio_digest(urls, voice, provider): Executes the full audio pipeline and returns the absolute file path to the saved MP3.

Architecture

The project follows a decoupled architecture to support multiple interfaces:

• src/core/: Interface-agnostic business logic for YouTube extraction, AI summarization, and TTS.
• src/interfaces/cli/: Command-line interface implementation.
• src/interfaces/api/: FastAPI backend implementation.
• src/interfaces/mcp/: Model Context Protocol server implementation.