MCP Leave Management Server

A simple Model Context Protocol (MCP) server that exposes employee leave-management operations (apply, cancel, check balance, view history) as tools an AI client (like Claude Desktop, Claude Code, or the MCP Inspector) can call directly.

Built as a learning project for understanding how MCP servers work — how tools are defined, registered, and invoked by an LLM client — before extending it into a real backend (e.g. with persistent storage).

Features

• Add new employees with a default leave balance

• Check an employee's remaining leave balance

• Apply for leave (with automatic balance deduction and history logging)

• View an employee's full leave history

• Cancel a previously applied leave (restores balance)

• List all employees in the system

Note: Data is stored in-memory for this version — it resets every time the server restarts. Persistence (SQLite/Postgres) is a planned next step.

Related MCP server: HR Leave Management MCP Server

Tech Stack

• Python 3.10+

• mcp (Python MCP SDK, FastMCP)

• uv for environment and dependency management

Project Structure

mcp-leave-management/
├── leave_server.py      # MCP server with all tool definitions
├── pyproject.toml       # Project metadata & dependencies (created by uv)
├── README.md
└── .gitignore

Setup

1. Clone the repository

git clone <your-repo-url>
cd mcp-leave-management

2. Install dependencies with uv

uv init --no-workspace
uv add "mcp[cli]"

3. Run the server with MCP Inspector (for manual testing)

uv run mcp dev leave_server.py

This opens the MCP Inspector in your browser, where you can connect to the server and call each tool manually with test inputs.

Available Tools

Default seed data

Example Usage (via MCP Inspector)

1. Call list_employees → confirms ["Chandu", "Sekhar"]

2. Call get_leave_balance with emp_name="Chandu" → returns 10

3. Call apply_leave with emp_name="Chandu", days=2, reason="Personal" → returns a confirmation message and deducts the balance

4. Call get_leave_balance again for Chandu → returns 8

5. Call get_leave_history for Chandu → shows the applied leave entry with its id

6. Call cancel_leave with that leave_id → restores the balance back to 10

Connecting to Claude Desktop / Claude Code

Once verified in the Inspector, this server can be registered as an MCP tool in a client's configuration (e.g. claude_desktop_config.json), pointing to the uv run command for leave_server.py.

Roadmap / Next Steps

• Replace in-memory dict with persistent storage (SQLite or Postgres)

• Input validation (e.g. reject negative/zero days in apply_leave)

• Role-based access (manager approval flow for leave requests)

• Unit tests for each tool

• Docker support for deployment

License

MIT (or update as preferred)