convez.ai
MCP ServerAutomaçãoAvançado
HendoCode

Contact Center AI MCP Server

HendoCode/contact-center-ai

RAG + MCP para inteligência em contact centers, agnóstico a providers

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que expõe um pipeline RAG para análise de transcrições de chamadas, gravações e dados de satisfação de clientes. Integra LangChain, pgvector e OpenAI em um servidor reutilizável que qualquer cliente compatível com MCP (Claude, Gemini, etc.) pode consumir sem reconstruir a camada de dados.

Instalação

$npx convez-ai HendoCode/contact-center-ai

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 4 dias
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Análise de transcrições de chamadas com busca semântica
  • Consultas de satisfação do cliente (CSAT) por período e filtros
  • Identificação de padrões e reclamações comuns em atendimentos
  • Acesso empresarial compartilhado via autenticação Azure Entra

Veredito

Solução robusta e production-ready para supervisores de contact center consultarem dados de chamadas via linguagem natural. Destaca-se pela arquitetura agnóstica a providers (LLM e embeddings) e autenticação corporativa integrada; demanda experiência em Python, Docker e infraestrutura cloud.

README

A tradução para o português deste README ainda está na fila.

contact-center-ai

A RAG pipeline and MCP server that gives contact center supervisors natural language access to call recordings, transcripts, and CSAT survey data.

Built to be provider-agnostic (swap OpenAI for Anthropic or Gemini via LangChain) and reusable — the MCP server is designed as a shared enterprise tool, not tied to a single team or use case.

What it does

  • Ingests call transcripts and CSAT survey results into a pgvector store
  • Exposes a RAG-powered query interface: ask questions like "what were the most common complaints last week?" or "show me calls where the member mentioned fraud"
  • Serves these capabilities as MCP tools so any GenAI client (Claude, OpenAI, Gemini) can consume them without rebuilding the data layer

Architecture

Call recordings / transcripts / CSAT surveys
         │
         ▼
    S3 (staging)
         │
         ▼
  Ingestion pipeline
  (LangChain + OpenAI embeddings)
         │
         ▼
   pgvector (PostgreSQL)
         │
         ▼
   MCP Server (Python)
   ├── search_transcripts(query)
   ├── get_call_summary(call_id)
   └── query_csat(date_range, filters)
         │
         ▼
  Any MCP-compatible GenAI client
  (Claude Desktop, custom app, etc.)

Auth: Microsoft Entra via Azure App Service Easy Auth for MCP — governs which teams and identities can access the server in production.

Stack

  • Python 3.12 (managed by uv)
  • LangChain — orchestration and LLM abstraction
  • pgvector — vector similarity search inside PostgreSQL
  • OpenAI — embeddings and completions (default; swappable)
  • Ollama — local embeddings and completions, no API key required (optional)
  • Python MCP SDK — MCP server implementation
  • Terraform — Azure infrastructure (App Service + PostgreSQL)
  • Microsoft Entra — authentication (production)

Quick start

Prerequisites

  • Docker and Docker Compose
  • uv (manages Python 3.12 automatically)
  • OpenAI API key — or use Ollama locally (see Swapping LLM providers)

Run locally

# Clone the repo
git clone https://github.com/yourusername/contact-center-ai
cd contact-center-ai

# Copy and configure environment
cp .env.example .env
# Edit .env — set OPENAI_API_KEY, or set LLM_PROVIDER=ollama

# Start PostgreSQL with pgvector (and Ollama, if using it)
docker compose up -d

# Create virtualenv and install dependencies
uv venv --python 3.12
uv sync --extra dev

# Generate synthetic call data
python data/synthetic/generate_data.py

# Ingest data into vector store
python -m rag.pipeline --ingest

# Start the MCP server
python -m mcp.server

Connecting a client

Once the MCP server is running locally, add it to your Claude Desktop config:

{
  "mcpServers": {
    "contact-center": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/contact-center-ai", "python", "-m", "mcp.server"]
    }
  }
}

Swapping LLM providers

Set LLM_PROVIDER in your .env to switch providers — no code changes needed.

Ollama (local, no API key required)

Ollama runs as a Docker Compose service — docker compose up -d starts it automatically.

# Pull models into the running Ollama container
docker compose exec ollama ollama pull llama3.2
docker compose exec ollama ollama pull nomic-embed-text

# In .env:
LLM_PROVIDER=ollama
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=llama3.2
OLLAMA_EMBEDDING_MODEL=nomic-embed-text

# Run the pipeline as normal
python -m rag.pipeline --ingest
python -m rag.pipeline --query "fraud disputes from last week"

OpenAI (default)

LLM_PROVIDER=openai   # or omit — openai is the default
OPENAI_API_KEY=sk-...

The pipeline is also designed so you can add other LangChain-supported providers (Anthropic, Google Vertex AI, etc.) by extending rag/embeddings.py and rag/pipeline.py.

Production deployment

See infra/terraform/ for Azure App Service + PostgreSQL infrastructure. Entra auth is configured at the App Service level — see infra/terraform/main.tf.

Notes

This repo uses synthetic data only. The real engagement this mirrors involves a confidential financial institution client. No real member data, call recordings, or CSAT results are included anywhere in this repository.

Ver o original no GitHub

Bom para

Relacionados