Secure Local MCP Agent (Engineering Edition)

Đọc bản tiếng Việt

A Software Engineer AI Agent system that is 100% offline, secure, and directly interacts with your personal computer via the MCP (Model Context Protocol).

🎯 The Problem Solved

Current Cloud AIs pose a risk of source code data leaks and cannot directly interact with your local operating system. This project builds a personal AI running 100% offline (Native) on Windows, capable of reading/analyzing your project's source code, as well as executing real system commands (like ping, ipconfig, dir) without being restricted by virtualization layers (Docker).

Related MCP server: Lexomni MCP

🏗️ System Architecture (Native Mode)

my_notes/ (Notes) & project_context/ (Source Code)
                        │
                        ▼
src/indexer.py ──── SQLite FTS5 ──── Scan & Index Code/Text
                        │
                        ▼
src/server.py ───── FastMCP ──────── 5 Tools (list, read, search, write, execute_command)
                        │
                        ▼ (stdio / JSON-RPC)
src/ui.py ───────── Streamlit ────── Visual Web Interface (Transparent Tool execution flow)
                        │
                        ▼
Ollama (Gemma4/Qwen) / Any Local LLM

✨ Highlight Features

🚀 Installation & Usage (1-Click)

The system is designed to run directly on Windows (Native) to fully utilize the operating system's capabilities and permissions.

Requirements:

• Python 3.10+ installed (with PATH configured).

• Ollama installed and running on the default port (http://localhost:11434).

How to run: Simply double-click the start.bat file. The script will automatically:

1. Create a virtual environment (venv).

2. Install the necessary dependencies (pip install).

3. Open the Web UI app (Streamlit) in your browser at http://localhost:8501.

🛡️ Security Mechanisms

The system applies a robust protection layer against Directory Traversal Attacks (CWE-22): It utilizes resolve() and is_relative_to() to ensure that the read/write paths absolutely cannot escape the 2 permitted root directories (my_notes and project_context).

📁 Project Structure

secure-local-mcp/
├── my_notes/               # Personal notes repo (.md, .txt)
├── project_context/        # Project source code for AI analysis (.py, .js, .json,...)
├── src/
│   ├── server.py           # FastMCP Server (Tool declarations)
│   ├── indexer.py          # SQLite FTS5 Indexing Engine (Handles Text & Code)
│   └── ui.py               # Streamlit Web Chat Interface
├── start.bat               # Windows (Native) automated startup script
├── requirements.txt
└── README.md