Advanced AI Customer Support Agent with MCP

An intelligent customer support agent powered by Model Context Protocol (MCP) that integrates with CRMs, ticketing systems, communication tools, and databases to provide personalized, real-time assistance.

Features

• Real-time Data Access: Pull current order status, shipping details, account history from databases/APIs via MCP

• Context Awareness: Access previous conversations, tickets, emails, Slack threads through unified memory layer

• Action Capabilities: Update tickets in Zendesk, create new tickets, send notifications instantly

• Context Persistence: Maintain conversation history across channels

• Flexible LLM Support: Choose between Ollama (local, private), OpenAI (cloud, powerful), or Vertex AI/Gemini (Google Cloud)

• MCP Integrations: Secure connections to enterprise systems (CRMs, ERPs, helpdesk tools) in real-time

Related MCP server: CRMy

Architecture

customersupport/
├── agent/              # Main agent application
├── mcp_integrations/   # MCP server integrations
├── integrations/       # External system connectors
├── memory/             # Context persistence layer
├── config/             # Configuration files
├── utils/              # Utility functions
└── frontend/           # React UI for customer support chat

Quick Start

See SETUP.md for detailed installation and configuration instructions.

Basic Setup

0.Create developer accounts for Zendesk and Slack https://ekluvtech.com/2026/02/21/creating-developer-accounts-for-zendesk-and-slack/

1. Install dependencies:

  #windows
  python3  -m  venv custsupport
  .\custsupport\Scripts\activate

  #Mac
  python3.10 -m venv custsupport
  source custsupport/bin/activate

pip install -r requirements.txt

2. Create developer accounts

• Zendesk trial or dev/sandbox account

• Slack app in a development workspace

• Set up an order database(PostgreSQL/MySQL for production-like testing)

• Choose an LLM provider (OpenAI, Anthropic, Geminietc.) and get an API key

• export all the required environment varaibles

• export ZENDESK_SUBDOMAIN=*********@

• export ZENDESK_EMAIL=*********@gmail.com

• export ZENDESK_API_KEY=***************************

• export SLACK_TOKEN=xoxp--10255901183920--9301425075be6f3170e95241c7

• export DATABASE_URL="postgresql://ordruser:Admin123@localhost/ordrmgmnt"

2. Configure the agent:

   cp config/config.example.yaml config/config.yaml
   # Edit config/config.yaml with your settings

3. Configure LLM Provider:

   The agent supports three LLM providers:

   Option A: Ollama (Local, Private)

   • Start Ollama (if not already running):

     ollama pull llama2

   • Configure in config/config.yaml:

     llm:
       provider: ollama

   Option B: OpenAI (Cloud, Powerful)

   • Set your OpenAI API key:

     export OPENAI_API_KEY=your-api-key-here

   • Configure in config/config.yaml:

     llm:
       provider: openai
     openai:
       api_key: "${OPENAI_API_KEY}"
       model: gpt-4

   Option C: Vertex AI / Gemini (Google Cloud)

   • Option 1: Service Account JSON (Recommended)

     • Create a service account in Google Cloud Console

     • Download the JSON key file

     • Set credentials path:

       export GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account-key.json
       export GOOGLE_CLOUD_PROJECT=your-project-id

   • Install Vertex AI dependencies:

     pip install google-cloud-aiplatform

   • Configure in config/config.yaml:

     llm:
       provider: vertexai  # or "gemini"
     vertexai:
       credentials_path: "${GOOGLE_APPLICATION_CREDENTIALS}"  # For JSON file
       project_id: "${GOOGLE_CLOUD_PROJECT}"
       location: us-central1
       model: gemini-pro

   • See GOOGLE_CLOUD_SETUP.md for detailed setup instructions

   • See VERTEX_AI_TROUBLESHOOTING.md if you encounter rate limit errors (429)

4. Run MCP server:

  `python -m mcp_integrations.unified_server --http --port 8000` (running all the tools under unified server)

5. Start the API server:

   python -m agent.api

6. Modern chat interface for customers:

   python -m agent.api
   ``

#if you get any certificate issue while connection to slack pip install --upgrade certifi export SSL_CERT_FILE=$(python -m certifi)

Usage

The agent can be used via:

• Web UI (React): Modern chat interface for customers (see Frontend Setup)

• Interactive Chat: python -m examples.interactive_chat

• MCP Server: python -m mcp_integrations.unified_server --http --port 8000 (running all the tools under unified server)

• REST API: python -m agent.api (requires FastAPI/uvicorn)

• Programmatic: Import and use the SupportAgent class directly

Frontend Setup

1. Navigate to frontend directory:

   cd frontend

2. Install dependencies:

   npm install

3. Start the development server:

   npm run dev

   The React app will open at http://localhost:3000

4. Make sure the MCP Server is running:

   # In a separate terminal
    python -m mcp_integrations.unified_server --http --port 8000

   The MCP Server runs on http://localhost:8000/ by default. and you can verify the tools available by querying http://localhost:8000/tools

5. Make sure the backend API is running:

   # In a separate terminal
   python -m agent.api

   The API runs on http://localhost:8100 by default.

6. Start chatting! The UI will automatically connect to the backend API.

For more details, see frontend/README.md.

Demo Data

To populate Zendesk, Salesforce, Slack, and your database with sample data for testing:

1. See demo_data/README.md for detailed instructions

2. Set up environment variables for each service

3. Run the population scripts:

   # Populate all services
   python demo_data/populate_all.py
   
   # Or populate individually:
   python demo_data/populate_database.py
   python demo_data/populate_zendesk.py
   python demo_data/populate_salesforce.py
   python demo_data/populate_slack.py

This will create sample tickets, orders, cases, and messages that you can use to test the agent's capabilities.

Privacy & Security

• Ollama (Local): All data processed locally, no data leaves customer premises

• OpenAI (Cloud): Data is sent to OpenAI's API - review OpenAI's privacy policy

• Vertex AI (Google Cloud): Data is sent to Google Cloud - review Google Cloud's privacy policy

• Secure MCP connections with authentication

• Customer identity verification before accessing sensitive data

• Configurable encryption and data retention policies

License

MIT