What is Agent Mcp Gemini Demo

agent-mcp-gemini-demo is a demonstration project that utilizes the Google ADK with a local MCP server to process YouTube videos, extract transcripts, and validate factual claims using FastAPI and Python.

Use cases

Use cases include verifying claims made in YouTube videos, educational content analysis, and enhancing customer service interactions by providing accurate information from video sources.

How to use

To use agent-mcp-gemini-demo, set up the FastAPI backend and connect it to the MCP server. Input a YouTube video URL or ID to initiate the agent pipeline, which will extract transcripts and perform fact-checking.

Key features

Key features include YouTube video processing, transcript fetching using MCP, claim extraction via a Gemini-powered LLM Agent, search planning for claims, and fact-checking through simulated Google Search.

Where to use

agent-mcp-gemini-demo can be used in fields such as education, media analysis, and research, where video content needs to be analyzed for factual accuracy.

Code Agent Gemini - Demo

This project is intended to showcase how to leverge the google-adk with a local MCP-server. We do this by attaching the MCP server created to the ADK adgent via MCPToolSet.

The agent pipeline uses sequential agents and loop agents to ingest YouTube video (via its URL or ID) through a local MCP-server, extracts its transcripts, identifies factual claims against web search results and validates those claims using a built-in google_search.

It features a FastAPI backend that processes YouTube video URLs, extracts transcripts, identifies factual claims, plans search queries, performs fact-checking using simulated Google Search, and presents a report.

Note:: While the FastAPI server can be used to service the agent endpoint, we are mainly going to leverge the build in adk web command

Demo

Overview of Agentic Capabilities:

• YouTube Video Processing: Accepts a YouTube video URL
• Transcript Fetching: Retrieves video transcripts using an MCP (Model Context Protocol) server powered by youtube_transcript_api
• Claim Extraction: Uses a Gemini-powered LLM Agent to identify key factual claims from the transcript
• Search Planning: Another LLM Agent devises Google Search queries for each claim
• Fact-Checking Loop: An ADK LoopAgent iterates through claims
  • Dequeues a claim.
  • An LLM worker agent (simulates) uses google_search and determines if the claim is True, False, or Unverified
  • Collects ‘verdicts’
• Sequential Orchestration: All steps are managed by an ADK SequentialAgent
• FastAPI Backend: Exposes an API endpoint to trigger the pipeline
• Dockerized: Includes Dockerfile and docker-compose for containerized deployment
• Installable Backend Package: The fastapi_build backend module is structured as an installable Python package

Directory Structure

code-agent-gemini/
├─ .devcontainer/
│  └─ devcontainer.json
├─ .github/
│  └─ workflows/
│     └─ python.yaml
├─ backend/
│  ├─ fastapi_build/
│  │  ├─ agents/
│  │  │  ├─ __init__.py
│  │  │  └─ youtube_processing_agents.py
│  │  ├─ core/
│  │  │  ├─ __init__.py
│  │  │  └─ config.py
│  │  ├─ mcp_servers/
│  │  │  ├─ __init__.py
│  │  │  └─ youtube_transcript_mcp_server.py
│  │  ├─ tools/
│  │  │  └─ __init__.py
│  │  ├─ __init__.py
│  │  ├─ agent.py
│  │  └─ main.py
│  ├─ fastapi_build.egg-info/
│  ├─ tests/
│  │  ├─ __init__.py
│  │  └─ test_youtube_processing_agents.py
│  ├─ Dockerfile
│  ├─ pyproject.toml
│  ├─ README.md
│  ├─ requirements-dev.txt
│  └─ requirements.txt
├─ frontend/
├─ video/
│  ├─ demo-video.mp4
│  └─ thunbnail.png
├─ .env.example
├─ .gitignore
├─ .pre-commit-config.yaml
├─ docker-compose.yml
├─ LICENSE.md
├─ pyproject.toml
└─ README.md

Prerequisites

• Python (version 3.11+ recommended, see backend/Dockerfile for version used in container)
• Google API Key for Generative AI (Gemini)
  • Obtain from Google AI Studio or Google Cloud Console
• gcloud CLI (if using Application Default Credentials locally, or for Vertex AI)
• Docker and Docker Compose (for containerized deployment)

Setup and Installation (Local)

1. Clone the Repository:

   git clone https://github.com/abdulzedan/agent-mcp-gemini-demo.git
   cd code-agent-gemini
   

2. Set up Environment Variables:
   Copy .env.example to .env in the project root and fill in your GOOGLE_API_KEY and other relevant details:

   cp .env.example .env
   # Edit .env with your credentials
   

   Example .env content:

   GOOGLE_API_KEY="YOUR_GOOGLE_API_KEY"
   
   GOOGLE_CLOUD_PROJECT="your-gcp-project-id" # Optional, if not using Vertex AI for Gemini this can be a placeholder
   GOOGLE_CLOUD_LOCATION="us-central1"      # Optional, same as above
   LOG_LEVEL="INFO"
   GOOGLE_GENAI_USE_VERTEXAI="false" # Set to true if using Vertex AI Gemini models
   

3. Backend Setup:
   Create a Python virtual environment:

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate
   

   Navigate to the backend directory:

   cd backend
   

   Install dependencies and the fastapi_build package in editable mode:

   pip install --upgrade pip
   pip install -r requirements.txt
   pip install -r requirements-dev.txt # For testing and development tools
   pip install -e . # Installs fastapi_build from pyproject.toml
   

4. Google Authentication (Local Development):
   If GOOGLE_GENAI_USE_VERTEXAI is true or if your agents/tools need broader Google Cloud access, ensure you have Application Default Credentials:

   gcloud auth application-default login
   

   For Gemini API directly with an API key (GOOGLE_GENAI_USE_VERTEXAI="false"), this step might not be strictly necessary if the key is the only auth needed by google-generativeai.

Running the Application (Local)

1. Exposing the ADK Web tool to see monitor agent flows:
   From the backend directory (with the virtual environment activated):

   adk web
   

   The adk web UI will load and will be available at http://localhost:8000.

Note:: When you open the adk web UI, you will need to select the “fastapi_build” on the agents

2. MCP Server (youtube_transcript_mcp_server.py):
   This server is started on-demand by the TranscriptFetcherAgent using StdioServerParameters. You don’t need to run it separately
   For standalone testing of the MCP script:

   # From the backend directory
   python fastapi_build/mcp_servers/youtube_transcript_mcp_server.py
   

Running with Docker

1. Ensure your .env file is created in the project root with your GOOGLE_API_KEY.

2. Build and Run using Docker Compose:
   From the project root directory (code-agent-gemini/):

   docker-compose up --build
   

   To run in detached mode:

   docker-compose up --build -d
   

3. To Stop Docker Compose:

   docker-compose down
   

API Endpoints

Note:: This is if you are intending to run the FastAPI server. if that is the case,
please head to the OpenAPI through appending /docs to the http://localhost:8000

• POST /process-video/:

  • Processes a YouTube video.
  • Request Body (JSON):

    {
      "video_url": "[https://www.youtube.com/watch?v=your_video_id](https://www.youtube.com/watch?v=your_video_id)",
      "user_id": "optional_user_identifier",
      "session_id": "optional_session_to_continue"
    }
    

  • Response Body (JSON):

    {
      "session_id": "string",
      "summary": "string | null",
      "fact_check_report": "string | null",
      "full_agent_output": "string",
      "error": "string | null",
      "grounding_html_content": "string | null"
    }
    

• GET /:

  • Welcome message.

Running Tests

Unit tests are located in backend/tests/.

1. Ensure development dependencies are installed (see Backend Setup).
2. From the backend directory (with virtual environment activated):

   pytest
   

   Or from the project root:

   pytest backend/tests
   

Google Agent Development Kit (ADK)

This project heavily utilizes the Google Agent Development Kit (ADK) to structure and run the AI agents. Key ADK components used:

• BaseAgent, LlmAgent, SequentialAgent, LoopAgent
• Runner for executing agents.
• InMemorySessionService for session management.
• FunctionTool and MCPToolset for integrating external capabilities like transcript fetching and Google Search.

Pre-commit Hooks

This project uses pre-commit for code quality. To set it up:

pip install pre-commit
pre-commit install