Brokerage MCP server

I wanted to use Claude and Codex for portfolio reviews without handing a language model the keys to a brokerage account. The Interactive Brokers API is large, and exposing it one endpoint per tool gives an agent too many ways to do the wrong thing. So I built a small MCP server with an opinionated tool surface instead.

Small tools, clear boundaries

Twelve tools map to workflows rather than endpoints. portfolio returns the account context an agent needs before suggesting anything. position_size does advisory risk sizing from account value, stop distance and exposure. submit_order returns a dry-run bracket preview by default.

Multi-step judgement lives in the client’s prompts and skills. The server’s job is to enforce the boundaries.

Safety model

• Demo mode is the default, with synthetic account, market and option data, so anyone can try it without broker credentials
• Orders are dry-run previews unless live mode and live trading are both explicitly switched on
• Every mutating action is written to a local SQLite audit log
• Live submit, cancel and modify are designed to sit behind a human approval step

Try it

It runs in demo mode out of the box:

uv sync
AGENTIC_BROKERAGE_MCP_BROKER_MODE=demo uv run agentic-brokerage-mcp

The public repo is extracted from a larger private project, without the agent daemon, backtesting code and personal settings.

Then the official one arrived

Interactive Brokers has since released an official connector, which covers most of what I built this for. The parts I’d still carry into any agent that can move money are the ones this project was really about: preview first, keep an audit trail, and keep the tool surface narrow.