Local Agentic AI System — MCP + RAG + Skills

A modular, local-first AI agent system combining MCP (Model Context Protocol), RAG (Retrieval-Augmented Generation), and Skills (execution layer) into a unified architecture. Runs entirely on your machine with no cloud dependency.

Optimized for NVIDIA RTX 3050 (4GB VRAM), 16GB RAM, 512GB SSD, Intel CPU.

Architecture Overview

┌─────────────────────────────────────────────────────┐
│                   User Interface                      │
│          Streamlit / Gradio / FastAPI + React         │
├─────────────────────────────────────────────────────┤
│                    Orchestrator                       │
│     Routes queries → decides → executes + responds   │
├──────────┬──────────────────┬───────────────────────┤
│   RAG    │       MCP        │        Skills          │
│  Layer   │    Connection    │      Execution         │
│          │      Layer       │       Layer            │
├──────────┴──────────────────┴───────────────────────┤
│                   LLM (Ollama)                       │
│            Llama 3.1 8B / Mistral 7B / Qwen 2.5 7B   │
│                    RTX 3050 GPU                       │
└─────────────────────────────────────────────────────┘

Components

1. LLM Layer — src/llm/

• Ollama for local model serving
• Models: Llama 3.1 8B, Mistral 7B, Qwen 2.5 7B
• CUDA-accelerated inference on RTX 3050

2. RAG Layer — src/rag/

• ChromaDB as local vector store
• LangChain / LlamaIndex for orchestration
• Embeddings: all-MiniLM-L6-v2 or bge-base-en-v1.5
• Document ingestion pipeline (PDF, DOCX, TXT, HTML)

3. MCP Layer — src/mcp_servers/

• File system MCP server
• SQLite MCP server
• Custom industrial MCP servers (OPC UA, Modbus)
• Git MCP server for versioning

4. Skills Layer — src/skills/

• Python code execution (sandboxed)
• File read/write operations
• Data analysis (pandas, numpy, matplotlib)
• Report generation (Word, Excel, PDF)
• Database queries (SQLite, PostgreSQL)
• OPC UA / Modbus data acquisition

5. Interface — src/interface/

• Streamlit (rapid prototyping)
• Gradio (quick demos)
• Or FastAPI + React (production-grade)

Quick Start

Prerequisites

# 1. Install Ollama
# Download from https://ollama.com

# 2. Pull a model
ollama pull llama3.1:8b

# 3. Create Python virtual environment
python -m venv .venv
.venv\Scripts\activate

# 4. Install dependencies
pip install -r requirements.txt

# 5. Run the application
streamlit run src/main.py

Project Structure

mcp-rag-skills-agent/
├── config/                  # Configuration files
│   ├── settings.yaml        # Main configuration
│   └── skills_registry.yaml # Skills catalog
├── data/
│   ├── documents/           # Source documents for RAG
│   ├── chroma_db/           # Vector database storage
│   └── logs/                # Application logs
├── src/
│   ├── llm/                 # LLM integration
│   │   └── ollama_client.py
│   ├── rag/                 # RAG pipeline
│   │   ├── embeddings.py
│   │   ├── vector_store.py
│   │   └── retriever.py
│   ├── mcp_servers/         # MCP protocol servers
│   │   ├── file_server.py
│   │   ├── sqlite_server.py
│   │   └── mcp_client.py
│   ├── skills/              # Execution skills
│   │   ├── registry.py
│   │   ├── python_executor.py
│   │   └── data_analyzer.py
│   ├── interface/           # UI layer
│   │   ├── app.py           # Streamlit app
│   │   └── api.py           # FastAPI app
│   └── main.py              # Entry point
├── tests/
├── docs/
├── requirements.txt
├── README.md
├── ARCHITECTURE.md
└── CAHIER_DES_CHARGES.md

Configuration

Edit config/settings.yaml:

llm:
  provider: ollama
  model: llama3.1:8b
  temperature: 0.3
  max_tokens: 2048

rag:
  embedding_model: all-MiniLM-L6-v2
  chunk_size: 512
  chunk_overlap: 64
  top_k: 5

mcp:
  servers:
    - file_system
    - sqlite
    - git

skills:
  sandbox: true
  timeout: 30

Use Cases

Industrial Technical Assistant

• RAG on equipment manuals, procedures, standards
• Skills: OEE calculations, report generation, SQL queries
• MCP: local file access, database connectivity

Data Process Analyzer

• RAG on production history
• Skills: Python analysis (pandas), visualizations
• MCP: database connections, file I/O

Documentation Generator

• RAG on templates and existing docs
• Skills: Word/Excel/PDF creation
• MCP: Git integration for versioning

License

MIT