// GLOSSARY

AI trading agent

An LLM-driven agent (Claude, Codex, etc.) that researches markets and places orders through tools — typically an MCP server or REST API — rather than a human clicking a UI. Safe deployment requires server-side risk limits the agent cannot override.

An AI trading agent is a language model wired to trading tools: it can search markets, read order books, and submit orders the same way it calls any other tool. On prediction markets the fit is unusually good — contracts are questions about the world, and reasoning about the world is what LLMs do.

How agents connect

The common pattern is an MCP server exposing tools like search_markets and place_order, or a plain REST API the agent calls from generated code. Either way, the agent holds a scoped credential, not your exchange keys.

The safety problem

Agents are non-deterministic. A prompt bug or hallucinated ticker can turn into a real order, so the guardrails must live server-side — notional caps, slippage bounds, and a kill switch enforced by infrastructure the agent cannot talk its way past. Mithril's hosted MCP server takes this approach: every agent order is checked against your limits before it reaches a venue. See risk limits for AI trading agents.