TradePilot AI

Most "AI trading bots" are black boxes: you can't see why a trade was made, there's no hard risk control, and the LLM can place reckless orders. For real use — and for options especially you need transparency and guardrails. THE SOLUTION TradePilot AI separates judgment from control. A multi-agent pipeline produces a trade proposal WITH a plain-English rationale; a separate, deterministic (non-AI) risk engine has the final say on every order; and a timestamped audit log records the proposal, the risk verdict + reason, and the execution result. It trades equities (bracket orders) and options (long calls for bullish views, long puts for bearish views), all on Alpaca paper trading. HOW IT WORKS (pipeline) 1. Market Analyst agent — fetches price/volume/volatility via the Alpaca MCP server; summarizes trend, momentum, and levels. 2. News Sentiment agent — fetches recent news via MCP; scores sentiment from -1 to 1. 3. Decision agent — combines both + account state into a structured proposal with a reasoning field for explainability. 4. Risk engine (deterministic, pytest-covered) — enforces max position size, daily-loss limit, trades/day, concentration, duplicate protection, and required stop/target. Approves or rejects. 5. Execution via Alpaca CLI — stocks as bracket orders (entry + stop + target, legs anchored to the live price); options as single-leg long calls/puts, near-the-money, sized to the risk budget. 6. Audit trail in Postgres, surfaced on a Next.js dashboard. Optional human-in-the-loop approval queue. HOW WE USED ALPACA - Trading API: account, orders, and market + options data (REST). - MCP server: the AI agents fetch data through alpacahq/alpaca-mcp-server, restricted to a READ-ONLY tool allowlist (the agents can never see order-placement tools). - CLI: order execution (alpaca order submit), account, and options chain.