This is a personal trading platform I’ve been building for a while. Recently I’ve been handing more of it to AI agents. Trading is a good place to test that, because a mistake costs real money right away. It all runs on a paper account.
How it works
- There’s an MCP server with 32 tools: quotes, options chains, read-only SQL over the local database, pattern scans (cup and handle, bull flag, double bottom, breakouts), and order entry.
- Events like fills, rejections, risk-limit breaches, and dropped connections go to an agent inbox, and they can wake an agent. There’s a cooldown and only one run in flight at a time.
- Order guardrails the agent can’t talk its way around: a symbol allow-list, a max number of orders per day, and a max dollar amount per order. If there’s no price, the order is rejected.
- Every order the agent tries, accepted or rejected, goes into a journal, and the agent has to give its reasoning. A blank reason means no order.
Why it’s here
The guardrail-and-journal pattern isn’t really about trading. It’s what I’d put around any agent that touches a real system: purchasing, scheduling, customer email.