Skincare

You install and run a Python-based MCP server that bridges a skincare ingredient recommendation engine to Claude. It exposes two practical tools for end users: finding similar products by ingredient lists and checking ingredients for irritants, helping you tailor recommendations to sensitive skin needs.

How to use

You connect an MCP client to the server and start issuing tool calls. The server provides two tools you can invoke in your Claude workflow:

• find_similar_products: provide a product’s ingredient list to obtain the top 5 most similar products by cosine similarity, using a TF-IDF representation with bigram support.

• check_red_flags: scan a product’s ingredients for known irritants and receive flags indicating potential concerns for sensitive skin users.

How to install

Prerequisites you need before setup: Python 3.11 or newer and the uv package manager.

Step 1: Clone the project and prepare the environment.

git clone https://github.com/pserein/skincare-mcp.git
cd skincare-mcp
uv sync

# Download the dataset and place it in the project root
# cosmetic_p.csv is available from Kaggle: https://www.kaggle.com/datasets/eward96/skincare-products-clean-dataset

.venv/bin/python generate_user_history.py

Step 2: Configure the MCP client to connect to this server. Use the runtime command shown below to start the MCP server from your environment.

{
  "mcpServers": {
    "skincare-recommender": {
      "command": "/path/to/.venv/bin/python",
      "args": ["/path/to/skincare-mcp/mcp_server.py"]
    }
  }
}

Additional notes on usage and startup

The server is a stdio MCP configuration, so you start it using the Python interpreter from your virtual environment followed by the MCP server entry point. The setup steps above generate a synthetic user history dataset used for offline reinforcement learning experiments, and a dataset of skincare products to support similarity queries.

Configuration and notes

Prerequisites include Python 3.11+ and the uv tool for managing MCP configurations. You will generate synthetic user history with generate_user_history.py, which creates a structured State-Action-Reward dataset compatible with offline RL methods.

Troubleshooting and tips

If you encounter issues starting the server, verify that the Python virtual environment is activated and the path to mcp_server.py is correct. Ensure the cosmetic_p.csv dataset exists at the project root and that generate_user_history.py has completed without errors before attempting to query the server.

Available tools

find_similar_products

Vectorizes a product’s ingredient list using TF-IDF and returns the top 5 most similar products by cosine similarity, with bigram support.

check_red_flags

Scans a product’s ingredients for known irritants and flags potential concerns for sensitive skin users.