MCP (Model Context Protocol) lets AI assistants call external tools. I built a remote MCP server called Helium that gives any MCP-compatible AI assistant access to real-time financial intelligence — market data, ML-powered options pricing, news bias analysis, and more.
The interesting part isn't the financial data itself (there are plenty of market APIs). It's what happens when you combine structured financial data with an LLM's reasoning.
Setup (30 seconds)
Add one line to your AI assistant's MCP config:
Cursor / Windsurf:
{"mcpServers":{"helium":{"url":"https://heliumtrades.com/mcp"}}}
Claude Desktop:
{"mcpServers":{"helium":{"command":"npx","args":["mcp-remote","https://heliumtrades.com/mcp"]}}}
No API key. Nothing to install. Free tier.
What it does
Helium exposes 10 tools through MCP:
1. Market Intelligence (get_ticker)
Ask "What's the outlook for NVDA?" and get:
- AI-generated bull and bear cases
- 5 probability-weighted scenarios (e.g., 38% chance of mean-reversion, 25% upside on AI headlines, 10% tail risk on export shock)
- Each scenario includes falsifiability criteria — what would prove it wrong
2. ML Options Pricing (get_top_trading_strategies)
The model computes independent fair values for every listed options contract. For each ticker, it returns:
- Strategies ranked by expected value
- Backtested win rates (e.g., short vol calls on AAPL: 61% win rate, avg +$8.40/trade over 39 historical trades)
- Full Greeks for every contract
A single call returns ~355KB of structured data.
3. Balanced News Synthesis (search_balanced_news)
Aggregates 3.2M+ articles from 5,000+ sources. Instead of one take, it shows where sources agree vs. diverge on any topic.
4. Multi-Dimensional Bias Scoring (get_all_source_biases)
Scores news sources across 15+ dimensions — not just left/right:
- Prescriptiveness: Does the outlet tell you what to think, or just report?
- Sensationalism: Framing intensity
- Fearful framing: How much fear-based language is used
- Integrity: Factual rigor
- Plus: dovish/hawkish, libertarian/authoritarian, and more
5. Historical Options Data (get_historical_options_data)
Full historical chains with ML pricing baked in. A single SPY request returns ~30MB of every contract with the model's theoretical value.
Why MCP matters for this
The key insight is that MCP eliminates the build step. Instead of building a custom financial app, you add one config line and ask questions in natural language. The AI handles parsing 355KB of structured options data and pulling out what's relevant.
This pattern — domain-specific intelligence delivered through MCP — is how I think a lot of specialized AI tools will work going forward.
Links
Happy to answer questions about the implementation or the MCP protocol in general.
Top comments (0)