MCP Server

🧩 Main modules

1. Main MCP server - the core of the system, performing the following functions:

• Registry of connected servers

• Routing requests between servers

• Monitoring server status

• Aggregation of information about available tools

2. Specialized servers (connect to the main one):

• Embedding server - working with vector representations of text

• PDF extract server - conversion and extraction from PDF to Markdown

• Reranker server - ranking text data

• Qdrant server - managing vector collections

• PostgreSQL server - executing SQL queries and schema inspection

• LLM server - generating and streaming LLM responses, list of models

• MarkUp server - text/file markup using markup service methods

• Transcribe server - audio loading, status and transcription result

⚙️ Available tools

Note:

To add a new FastMCP server, you need to import it in the main_server.py file and place it in the MCP_SERVERS array, after which the main methods of main_server.py will have access to it.

Also a necessary requirement for FastMCP servers is the presence of the health_check method to check the state.

📡 Main server methods

Setting up the environment

Create a .env file in the root of the project and put the following environment variables in it (the list corresponds to the use in the code):

Start the main server

Installation dependencies and creating a virtual environment

Before starting the server, it is recommended to create a virtual environment and install all dependencies from requirements.txt. Run the following commands in the terminal:

1. Creating a virtual environment

2. Activating the environment

3. Installing dependencies

After setting up the environment, the server is started with the command

Running in Docker

1. Build the image:

2. Run the container, passing .env as environment variables:

Configuring the server connection in Cursor

1. Run the server using the command above

2. Open the settings

3. Add the MCP server configuration:

Local MCP server: proxy_mcp_server

• What is it: proxy MCP server that connects to the main registry (main_server) and forwards its methods, and provides a high-level pipeline for pre-preparing data for RAG.

• Where is it: proxy_mcp_server/proxy_mcp_server.py

• Available tools:

• get_server_and_tools() — get a list of servers and their tools from the registry

• router(server_name: str, tool_name: str, params: dict) — universal call router

• preprocessing_data_for_rag(file_paths: List[str]) -> str — prepare PDF/texts and create a collection in Qdrant; returns the collection name

• health_check_servers() — check if all services are available

Requirements:

• main_server is running and accessible via URL (e.g. http://localhost:8000/mcp/).

• Valid API key MAIN_SERVER_API_KEY (must match Authorization header in proxy_mcp_server.py).

• Update url and headers.Authorization in config object inside proxy_mcp_server/proxy_mcp_server.py if necessary.

Connection in Cursor (example):

Launch from terminal:

RAG inference: interactive launch

• What is this: console assistant for asking questions to a collection of documents in Qdrant with additional ranking and generation of LLM response.

• Location: rag_inference/RAG workflow.py

• Preliminary environment variables: QDRANT_URL, QDRANT_API_KEY, RERANK_URL, RERANK_MODEL, LLM_SERVICE_CHAT_COMPLETIONS_URL, LLM_SERVICE_API_KEY, LLM_SERVICE_MODEL, EMBEDDING_URL, EMBEDDING_MODEL are used (described above in the settings section).

Run (Windows PowerShell):

Where <collection_name> is the name of the collection in Qdrant. It is convenient to get it in advance by calling the preprocessing_data_for_rag tool from proxy_mcp_server and passing a list of files to index; the method will return the name of the created collection.

Example: