Helmsman | hermes agents

Helmsman is a multi-agent AI options trading system built on Alpaca's infrastructure. BackOffice, Research, Trader, and Bot are orchestrated by an APScheduler engine on a precise market-hours schedule, with a real-time STDIN interface for portfolio state, positions, blotter, and on-demand LLM queries. The core design principle is a hard separation between AI and execution. LLM agents reason about what to trade — Research proposes strategies each morning, Trader monitors risk throughout the day — but a deterministic Python Bot decides whether an order is allowed and places it. The Bot never calls an LLM. Every rule that can move money is hardcoded, auditable, and readable in one sitting. If the inference provider goes down, the Bot keeps running. Research fetches live Greeks directly from Alpaca's option chain via OptionChainRequest for strike selection. For 1DTE iron condors on SPY and QQQ, short strikes are placed just outside the market-implied expected move, so condor width automatically adjusts to current vol rather than a fixed assumption. Short put spreads target 0.20 delta at 7-10 DTE across eight underlyings — AAPL, MSFT, GOOGL, AMZN, META, NVDA, SPY, QQQ — exiting at 50% profit or a DTE threshold. Each evening, Research and Trader run a reflection pass over the episodic log — every trade, rejection, and risk event from the day. Repeating patterns are distilled into skill documents that agents consult the following morning, so the system learns from its own trading history over time. The operator interacts through the running Helmsman terminal — portfolio view, spread blotter, win rate table, upcoming jobs, strategy registry — with no dashboards or web UI required. The Alpaca MCP server is integrated as an advisory layer for open-ended natural language queries against live account data, without touching the execution path.