💰 Expense Tracker using MCP (FastMCP + LangChain + Ollama)- Sample Project for understanding MCP
This project demonstrates a simple end-to-end MCP (Model Context Protocol) example where:
-
A FastMCP server exposes tools to manage expenses stored in SQLite
-
A LangChain client connects to the MCP server
-
An LLM (Llama 3.2 via Ollama) decides when to call tools
-
Natural language queries like
"Add my expense 500 to groceries" automatically trigger backend database operations
📌 Architecture Overview
User (CLI)
│
▼
LangChain Client (client.py)
│
│ MCP (stdio)
▼
FastMCP Server (main.py)
│
▼
SQLite Database (expenses.db)
Key Components
| Component | Description |
|---|---|
| FastMCP | Exposes database operations as tools |
| LangChain MCP Adapter | Connects LLM to MCP tools |
| Ollama (Llama 3.2:3b) | Interprets user intent and calls tools |
| SQLite | Persistent expense storage |
📂 Project Structure
.
├── main.py # FastMCP expense database server
├── client.py # LangChain MCP client with LLM
├── expenses.db # SQLite database (auto-created)
└── README.md
🚀 Features
- ✅ Add expenses using natural language
- ✅ View total expenses
- ✅ List all expenses
- ✅ Automatic tool selection by LLM
- ✅ Persistent storage using SQLite
- ✅ MCP-compliant architecture
🛠️ Tools Exposed by MCP Server
The FastMCP server exposes the following tools:
add_expense
Adds a new expense entry.
{
"amount": 500,
"category": "groceries",
"description": "weekly shopping"
}
get_total
Returns the total sum of all expenses.
get_all_expenses
Returns a list of all recorded expenses.
⚙️ Prerequisites
Make sure you have the following installed:
- Python 3.10+
- Ollama
- Llama 3.2 model
- uv (Python package runner)
ollama pull llama3.2:3b
📦 Install Dependencies
uv add fastmcp langchain langchain-mcp-adapters langchain-ollama
▶️ Running the Client
Update paths inside client.py:
"command": "/home/omkar/.local/bin/uv",
"args": [
"run",
"fastmcp",
"run",
"/full/path/to/main.py"
]
Then run:
uv run client.py
🧠 How It Works (Step-by-Step)
-
User enters a natural language query
-
LLM decides whether a tool is needed
-
If required:
- Tool name + arguments are generated
-
LangChain invokes MCP tool
-
Result is returned to LLM
-
LLM generates final user-friendly respons
Just tell me 👍