// GLOSSARY
MCP tool
A single callable capability an MCP server exposes to AI models — a name, a description the model reads, and a JSON schema for arguments. Good trading tools are narrow and validated: place_order with typed fields, not execute_anything.
An MCP tool is the unit of capability in the Model Context Protocol: a named function with a human-readable description and a JSON schema. The model reads the description to decide when to call it and the schema to know how.
Anatomy of a trading tool
A place_order tool might take market_id, side, size, and
limit_price as required, schema-validated fields. The description doubles
as the model's instructions — ambiguity there produces wrong calls, so tool
design is closer to API design plus prompt engineering than either alone.
Design principles for trading
- Narrow beats general. Four specific tools (
search_markets,place_order,cancel_order,get_portfolio— Mithril's set) are safer and more reliably used than one open-ended tool. - Validate server-side. The schema stops malformed calls; only the server can stop harmful well-formed ones — that's where notional caps live.
- Return structured, readable results the model can reason over in its next step.
A hands-on walkthrough of an agent using these tools: Claude + MCP for prediction markets.