AlphaPilot: Dual-MCP Trading Advisory on Binance Agent OS
How to run a deterministic risk layer beside Binance’s allowlisted Agent OS MCP—without handing your exchange keys to a custom backend.
What AlphaPilot is
AlphaPilot is an AI-assisted market operations agent for Binance. It was built for the Binance Agent OS Mini Hackathon (Track A) with a clear product line:
Find opportunities. Control risk. Optimize capital.
It is not a black-box autotrader and not a chatbot that “suggests” buys with no path to execution. It is a policy and workflow layer that:
- Scans the market with public Binance REST
- Scores candidates with deterministic strategies
- Validates every intent with a Risk Engine that never trusts AI-produced limits
- Publishes risk-checked TradePlans through a web UI and its own MCP server
- Lets an allowlisted AI client (Claude, Cursor, ChatGPT, Codex, VS Code, Grok Bot) place orders on Binance Agent OS MCP
- Records fills so AlphaPilot can monitor positions and surface exit logic
Why this architecture exists
Binance hosts an official MCP endpoint:
https://agent.binance.com/mcp/agentic
OAuth for that endpoint is restricted to a short client allowlist. Self-built backends—including AlphaPilot’s own process—hit:
The AI Agent you are using is not currently supported.
So AlphaPilot adopted Option A:
| Role | Who owns it |
|---|---|
| Market data & analysis | AlphaPilot (public REST) |
| Risk limits & TradePlans | AlphaPilot (DB + Risk Engine) |
| Account balances | Reported into AlphaPilot by the AI client |
| Order placement | Binance Agent OS MCP (inside allowlisted client) |
| Position monitoring | AlphaPilot (after ) |
AlphaPilot never stores a Binance API key or OAuth token.
What it was built for
- Operators who want AI assistance without giving a model exchange custody
- Hackathon / agent-native demos that must work with Binance’s real Agent OS constraints
- A dual-MCP workflow: one server for policy, one for execution
- Human-in-the-loop by default (
trading mode), with optional stricter or looser modes laterapproval_required
Core loop:
Public REST scan / chat analysis → Strategies + RiskEngine → TradePlan (proposed) → Human / AI approves → Order on Binance Agent OS MCP → record_fill → Position monitor + exit signals
