Agentic Trader

Shadow Mode

Governance rails for an AI trading agent.

Overview

Claude reads my brokerage data through Robinhood's own hosted Agentic Trading MCP; my Python package is the governance layer around it. A deterministic RiskGovernor evaluates every proposed trade against a fixed YAML risk policy and records each approve/reject decision, with reasons, to a local SQLite log. A reconciliation pipeline matches my CSV trade history against MCP order data using salted-hash IDs, and a standalone price-watch monitor pushes alerts without ever touching account data. The whole system runs in read-only shadow mode: the agent proposes, nothing is placed without my explicit approval, and the agent has placed zero trades.

What It Does

•Deterministic RiskGovernor: every proposed trade is checked against a fixed YAML risk policy (per-trade, daily, weekly, drawdown, and position-size limits)
•Every approve/reject decision recorded to a local SQLite log with its reasons
•CSV-to-MCP order reconciliation with salted-hash IDs; ambiguous matches route to review, never auto-resolved
•Standalone price-watch monitor that pushes alerts and never reads account data or trades
•No autonomous execution: read-only shadow mode, zero trades placed by the agent

Traction

•$29K+ across monitored accounts (Sep 2026)
•+40% (+$8.2K) on the primary account (Jul 29 - Sep 15, 2026)

Stack

PythonMCPClaudeSQLiteRobinhood