The Problem
Traders using Spark have access to rich, live ASX market data delivered via Dynamic Data Exchange (DDE) — a Windows protocol that works great with Excel but is completely invisible to modern AI tools.
There was no way to ask simple questions like “What is my P&L today?” or “Which stock is dragging my portfolio?” without switching between tools, running Excel macros, or pulling reports manually. The data existed — it just wasn’t accessible to AI in any useful way.
• DDE data is siloed inside Spark and only readable by legacy Windows applications like Excel
• Modern AI tools (including Claude) have no awareness of DDE — they cannot subscribe to feeds or read live prices
• No structured, queryable layer existed between Spark’s live data and any AI interface
• Traders had no way to run natural language analytics across positions, P&L, and orders
Our Solution
We built a four-layer bridge that connects Spark’s live DDE feed to Claude Desktop — so traders can ask questions in plain English and get real-time answers backed by live market data.
Everything runs locally on the trader’s Windows machine. No cloud services, no manual steps, no data leaving the machine. Once the system is running, asking Claude “Show my positions” returns a live portfolio table with P&L in under two seconds.
• A Python DDE listener polls Spark every 2 seconds for 6 ASX stocks and writes data to Redis and SQLite
• A FastAPI backend exposes 15 authenticated REST endpoints for quotes, positions, P&L, orders, and analytics
• An MCP server registers 18 tools with Claude Desktop — one for each type of trading question
• Claude Desktop picks the right tool based on the user’s question and returns a formatted, data-driven answer
The result: Market data rows captured, live P&L calculated in real time, and 30 days of order and P&L history — all queryable through a simple conversation with Claude.
Solution Architecture
The system is split into four clear layers. Each layer does one job and hands off to the next.
| # | Layer | File | What It Does |
| 1 | DDE Listener | dde_client.py | Connects to Spark via Windows DDE. Polls BHP, FMG, CBA, NAB, PLS, WOW every 2 seconds. Writes snapshots to Redis and inserts rows into SQLite. |
| 2 | Data Store | spark_data.db + Redis | Redis holds real-time snapshots (< 1ms reads). SQLite stores 72,090+ historical rows across 4 tables: quotes, positions, orders, pnl_daily. |
| 3 | API Layer | api/main.py | FastAPI with 15 GET endpoints. Protected by API key. Endpoints cover live quotes, price history, positions, P&L, orders, and analytics. Swagger docs at localhost:8000/docs. |
| 4 | MCP Server | spark_mcp.py | 18 Claude tools registered via claude_desktop_config.json. Runs over stdio. Claude calls these tools dynamically based on the user’s natural language question. |
The database has four tables: quotes (live DDE snapshots), positions (holdings with live P&L), orders (trade history), and pnl_daily (30-day P&L with win rate and drawdown stats).
Deliverables
• Windows DDE Listener —
• polls Spark live, writes to Redis and SQLite, includes rotating logs
• Redis Real-Time Cache —
• sub-millisecond price reads for the API layer
• SQLite Database —
• Data rows, 4 tables, properly indexed
• FastAPI Backend —
• 15 endpoints, API key auth, OpenAPI docs auto-generated at /docs
• MCP Server —
• 18 Claude tools covering quotes, positions, P&L, orders, and analytics
• Dummy Data Seeder —
• one command to populate positions, 30 days P&L, and ~60 orders for immediate testing
• Claude Desktop Config —
• ready-to-use claude_desktop_config.json wiring everything to Claude
• Project Documentation —
• full technical docs covering architecture, setup, API reference, and troubleshooting
Tech Stack
| Technology | Role |
| Python 3.11 + pywin32 | Core runtime and Windows DDE client — connects to Spark and reads live market data |
| Redis | In-memory store for real-time price snapshots with sub-millisecond read latency |
| SQLite | Persistent relational database for all historical quotes, positions, orders, and P&L data |
| FastAPI + uvicorn | REST API framework serving all 15 endpoints with async performance and auto OpenAPI docs |
| httpx | Async HTTP client used by the MCP server to call the FastAPI backend |
| MCP SDK (Anthropic) | Model Context Protocol — registers the 18 tools and manages stdio communication with Claude Desktop |
| Claude Desktop | AI front end — interprets natural language, selects the right tool, and formats the answer |
| Docker + docker-compose | Containerised Redis instance — single command startup |
| loguru + python-dotenv | Structured logging with rotation, and environment variable management via .env |
Business Impact
This project demonstrates something that hasn’t been done before: connecting a legacy Windows DDE trading platform directly to a modern AI assistant. The implications go beyond this one implementation.
For individual traders: Instead of switching between Spark, Excel, and manual reports, a trader can now ask one question and get a structured answer in under two seconds. Queries that used to take 10–15 minutes of spreadsheet work — like drawdown analysis or best/worst performer — are instant.
For trading firms and prop desks: The architecture is reusable. Any DDE-based platform can be bridged using the same pattern: DDE listener → data store → REST API → MCP server → Claude. Adding a new data source or a new type of analysis takes hours, not weeks. The system is also audit-ready — every query is logged, all data is timestamped, and all endpoints are authenticated.
For the broader fintech and AI industry: This proves that legacy IPC protocols like DDE can be modernised without replacing the underlying platform. It also establishes MCP as a viable production protocol for financial tooling — not just developer experiments. The barrier to AI adoption in traditional trading environments just got a lot lower.
Concretely: live data freshness is 2 seconds, time from question to answer is under 2 seconds, and extending the system with a new tool or endpoint takes a few hours. The POC is working — 72,090+ rows captured, all 18 Claude tools tested, positions and P&L queried live.
Demo Video
