This project is a technical demonstration of Model Context Protocol (MCP) integrations and agentic workflows using a broker's Paper Trading (simulated) environment. It is not financial advice, and the underlying logic is not a recommended trading strategy. Anyone attempting to implement automated trading systems with real capital assumes all financial and legal risk.
Can an agent operate a real protocol, safely?
When brokers began shipping official MCP servers, it opened a clean way to let an AI agent read structured market data and take defined actions through a sanctioned interface — instead of brittle scraping. The experiment: wire that up end to end, encode a strict rules engine, and run the whole thing against simulated (paper) trades to study agentic behavior with zero real-money exposure.
Local model → MCP → rules engine.
- Local Claude modelThe reasoning layer runs locally and speaks MCP — no bespoke API glue, just the standard protocol the broker publishes.
- Broker MCP server (paper environment)The official MCP server exposes market data and simulated order actions. Every action in this build targets the paper-trading sandbox.
- Signal-driven rules engineA written ruleset defines what the agent may do: which signals/triggers to watch, position sizing bounds, and the specific conditions under which a simulated buy or sell is allowed — the model acts only inside those guardrails.
Strip away the domain and this is a governed agent: a model given a real external system, a hard ruleset, and clear boundaries on what it's permitted to do. That is exactly the pattern behind safe business automation — an agent that can act on live systems only within explicit guardrails. This build is where I proved I can stand that up on a brand-new protocol the week it shipped.