Educhain based MCP Server (via Google Gemini)
This project devises an MCP server that handles various functions like: generating MCQs, flashcards, lesson-plans etc.,
Structure
Claude Desktop(Front end) <-> MCP Server <-> LLM (google Gemini)
Installation and Initialization
Recommended Python version: 3.10
Packages Manager: uv (Recommended) or pip
Step 0: Clone this repository
git clone Anudeep-CodeSpace/educhain_mcp_server.git
cd educhain_mcp_server
Step 1: Install uv
pip install uv # universal
brew install uv # mac os only
Step 2: Initialize project
uv init # initialize an already existing project
Step 3: Add required packages
# They contain all the required sub -packages in them
uv add "educhain" "mcp[cli]"
Step 4: Add your Google gemini api key
# inside .env file
GOOGLE_API_KEY=<Your Google api key without quotes>
Step 5: Debug your MCP server
uv run mcp dev main.py
It produces a tokenised proxy server at
http://localhost:6274/?PROXY_API_TOKEN=<proxy token>
Paste it and navigate to the link in a browser. Click "Connect" and you can debug your tools, resources and prompts in that site.
Step 6: Install Claude Desktop app
Install Claude Desktop app and login with your account(can be new).
Step 7: Add MCP server to Claude Desktop app
In the git repo folder run
# Adds the MCP Server to Claude Desktop client
uv run mcp install main.py
After that your claude_desktop_config.json should look like this:
{
"mcpServers": {
"Educhain - MCP server": {
"command": "absolute/path/to/uv",
"args": [
"run",
"--with",
"mcp[cli]",
"mcp",
"run",
"absolute/path/to/main/main.py"
]
}
}
}
Final step: Check for any discrepancies in the logs
All the logs are located at:
%APPDATA%\\Claude\\logs\\mcp.log # in windows
~/Library/Logs/Claude/mcp.log # in macOS
Metadata
get_info(about://info) resource lists out all the tools and resources provided by this server
Key Characteristics of this project
Modularity: Separating the server initialization (server.py), route handling (handlers.py), and the main entry point (main.py) makes the codebase clean, scalable, and easy to maintain.
Clear Schema Definitions: Use of Pydantic models in the schema directory. It ensures strong data validation, clear API contracts, and self-documenting code for requests and responses.
Dependency Injection: Passing the mcp server instance to handler functions (handle_resources(mcp)) is a good practice. It avoids circular dependencies and global state issues.
Use of Decorators: The @mcp.resource and @mcp.tool decorators provide a clean and declarative way to define the server's capabilities.
Known Issues
Claude Desktop client cannot direclty access Resource Templates (Beta stage)
For Example Claude Desktop client cannot access the generate_lessonplan resource(uri = lessonplan://{topic}) cannot be used directly as it is in beta stage and doesn't support dynamic resource uri's!!!
So Generate a lesson plan to teach algebra cannot invoke the generate_lessonplan tools!
Needs external LLM to generate content
Claude being a powerful llm cannot direclty generate content according to our tools and resources! (Hence I am using Gemini)
Key Contributors
Myself(Anudeep-CodeSpace), Chatgpt, Perplexity AI, Gemini(Free LLM)
Note
Node js(LTS) version is required for debugging pyenv is not recommended(That wasted a lot of time for me 😭)