mcp_server

marcusjhang/mcp_server

3.2

If you are the rightful owner of mcp_server and would like to certify it and/or have it hosted online, please leave a comment on the right or send an email to henry@mcphub.com.

The BookOps MCP + LLM Project is a setup for integrating a Model Context Protocol (MCP) server with Large Language Model (LLM) powered business logic to handle financial data queries.

Tools

  1. explain_pnl

    Provides explanations for profit and loss for a given book and date.

  2. audit_exposure

    Audits the main exposures for a given book.

πŸ“˜ BookOps MCP + LLM Project

This repository demonstrates a fast setup for:

  1. MCP Server using FastMCP and stdout/stdin transport.
  2. LLM-powered Business Logic within the MCP tool functions.
  3. Agent that handles user input and orchestrates LLM function-calling + MCP tool invocation.

πŸ“ File Structure

/ β”œβ”€β”€ agent.py # LLM agent that handles queries and calls MCP via subprocess β”œβ”€β”€ server.py # FastMCP server defining tools explain_pnl & audit_exposure β”œβ”€β”€ dummy_data/ β”‚ β”œβ”€β”€ trades.csv # Contains book_id, date, symbol, qty, side, price β”‚ └── positions.csv # Contains book_id, symbol, sector, qty └── README.md # This documentation

πŸ› οΈ Requirements

  • Python 3.10+ (recommended version: 3.12)
  • Install dependencies:
  pip install fastmcp openai pandas python-dotenv

β€’	Ensure your .env includes:

OPENAI_API_KEY=sk-xxx

βΈ»

πŸš€ Running the System (STDIO Mode)

  1. Start the MCP Server (in one terminal)

python server.py

This waits for client calls over stdin/stdout.

  1. Run the Agent (in another terminal)

python agent.py

  1. Example interaction:

Explain the PnL for book HF123 on 2024-06-01 LLM chose explain_pnl … { "book_id": "HF123", "date": "...", "pnl": -72500.0, "summary": "…" }

  1. Or ask:

What are the main exposures for HF123? LLM chose audit_exposure … { "book_id": "HF123", "top_exposures": [...], "summary": "..." }

βΈ»

πŸ”„ How it Works 1. Agent uses LLM function-calling to choose a tool & arguments from FUNCTIONS. 2. Agent launches the server subprocess and sends MCP stdio JSON-RPC: β€’ initialize β€’ notifications/initialized β€’ tools/call 3. Server processes the request: β€’ Executes business Python & makes internal LLM calls. 4. Server returns JSON response. 5. Agent captures and displays it.

🚧 Next Steps & Future Improvements

  1. Expand the Toolset
    Add more BookOps analytical toolsβ€”such as historical PnL comparison, VaR (Value at Risk) calculations, or trade anomaly detection.

  2. Add Session Memory and Context
    Implement lightweight session memory for the agent to track past queries (e.g., remember last book/date asked, or support follow-up questions).

  3. Switch to Persistent HTTP Deployment
    Move from launching the MCP server per query (STDIO mode) to a persistent HTTP-based deployment (via Flask, FastAPI, or fastmcp-http-proxy).