Investment Statement MCP Server

Parse and query investment statements from multiple institutions using natural language.

Features

• Multi-Institution Support: Currently supports Questrade (Interactive Brokers and Scotia coming soon)
• Hybrid Storage: SQLite for structured queries + LanceDB for semantic search
• Natural Language Queries: Ask questions like "What were my dividends in Q3?" or "Show my SPY holdings"
• Structured Tools: Programmatic access via MCP tools
• Archive Management: Keeps original PDFs and parsed JSON for reference

Installation

# Install dependencies
pip install -e ".[dev]"

Configuration

The server is configured via .mcp.json. Key environment variables:

• LANCEDB_PATH: Vector database location (default: ./data/lancedb)
• SQLITE_PATH: SQLite database location (default: ./data/statements.db)
• PDF_ARCHIVE_PATH: Original PDF storage (default: ./data/pdfs)
• JSON_ARCHIVE_PATH: Parsed JSON storage (default: ./data/json)
• MODEL_NAME: Embedding model (default: all-MiniLM-L6-v2)

Usage

Starting the Server

python -m src.server

Indexing Statements

# Via MCP tool
index_statement("/path/to/statement.pdf")

Querying Data

Natural Language Search:

search_statements("What were my total dividends in 2025?", limit=10)

Structured Queries:

get_holdings_by_symbol("SPY")
get_transactions_by_date("2025-01-01", "2025-09-30")
get_account_balance("51516162", "2025-09-30")

Resources

• statements://catalog - List all indexed statements
• statements://{statement_id} - Get specific statement
• statements://{statement_id}/summary - Account summary
• statements://{statement_id}/holdings - Holdings/positions
• statements://{statement_id}/transactions - Transaction history

Supported Statements

Questrade

• ✅ Account Summary (balances, FX rates, TFSA info)
• ✅ Holdings/Positions (securities with cost basis, P&L)
• ✅ Transactions (trades, dividends, deposits, withdrawals)

Interactive Brokers (Coming Soon)

• ⏳ Planned

Scotia (Coming Soon)

• ⏳ Planned

Development

# Run tests
pytest

# Run tests with coverage
pytest --cov=src --cov-report=html

# Format code
black src tests

# Lint code
ruff check src tests

Architecture

• FastMCP 2.0: MCP server framework
• pymupdf4llm: PDF parsing optimized for LLMs
• LanceDB: Vector database for semantic search
• SQLite: Structured metadata storage
• sentence-transformers: Embedding generation

License

MIT