# marketwingman

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Osmar Hernandez](https://cerebralvalley.ai/u/oherna25)
- **GitHub:** https://github.com/oherna25/trading_bot_project
- **Demo video:** https://youtu.be/8trJmAye_HQ
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4026

MarketWingMan is an AI options desk for your first trade and your thousandth, where the AI has to show its work.

The problem: retail traders get confident answers from chatbots ("SPY is going up, 90% sure"), lists from screeners, and exposure data behind paywalls, and nobody checks anybody. MarketWingMan turns that around. The engine does the math first, the models argue, and every model call is graded and written to a journal.

How it works:
- Its own math. It pulls live option chains and price bars from free-tier market APIs and computes dealer exposure (GEX, VEX, DEX, charm, vanna) locally from stored snapshots. That shows where price is likely to stall and where it could break loose, with no vendor feed.
- Models that get graded. A Muse model makes a directional call and has to commit. The deterministic engine makes its own call. When they disagree, the app flags a CONTRADICTION on screen and in the journal instead of burying it in a paragraph.
- For new traders: every trade card speaks plain English first (entry, target, max loss), sized to a saved risk profile. Everything is paper trading by default.
- For experienced traders: Muse Spark pitches the bull case and Muse Wonder argues the bear case on any card. A one-call stress test reprices every card under a shock (for example SPY -3%) before you size anything.
- MCP: MarketWingMan is an MCP server with 55 tools (morning brief, alerts in plain words, stress tests, debates, deployments, journal, and more). Any AI assistant or agent can use the same engine, numbers and journal as the app.
- A premium-selling agent with hard limits. You set paper capital and a per-trade cap. Muse Spark Open, running on Baseten, picks among deterministic candidates. Local code validates and sizes the trade, a risk gate approves or rejects it, and the agent manages the position to a target, a stop, or expiry.
- Receipts for everything. Every signal, intent, fill and model call (with tokens and cost) lands in one replayable journal. Live ordering ships disabled behind a toggle, every order still passes risk review, and the kill switch is always the user's.

The offline deterministic path always works: missing keys, budget or weights fall back to rule-based logic, and the full test suite (970+ tests) runs offline. It installs with pip on Windows, macOS and Linux, or with Docker Compose.

The model proposes. The engine decides. Never trade alone again.

## More from Meta Global AI Developer Hackathon

- [Expertium](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4023)
- [Crown & Fluff](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4024)
- [AI Symbiont](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4025)
- [Manuals-MCP](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4027)
- [Gr8 AI Solutions](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4028)
- [Zayas House](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4029)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4026. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt