AI Trading MCP is a Model Context Protocol (MCP) trading rig that gives Claude — or any MCP-compatible AI client — the ability to trade a live crypto portfolio through exactly three tools:Documentation Index
Fetch the complete documentation index at: https://mintlify.com/theonetrade/ai-trading-mcp/llms.txt
Use this file to discover all available pages before exploring further.
get_status, open_position, and close_position. The LLM is the signal source; the engine owns all risk math, TP/SL computation, and order execution. A GramJS scraper pipes a live Telegram channel — text posts and chart screenshots — directly into the get_status feed, making the AI aware of real-world signals without exposing any exchange credentials or control surfaces to the model.
Quick Start
Build the workspace, authorize Telegram, and launch your first paper trade in minutes.
MCP Tools Reference
Explore get_status, open_position, and close_position — the complete AI vocabulary.
Architecture Overview
Understand the two-process model, HTTP bridge, and how the engine guards the agent.
Live Trading Guide
Configure Binance API keys and switch from simulated fills to real spot orders.
Telegram Feed
How GramJS scrapes posts and chart images into MCP message blocks.
Configuration Reference
All environment variables, defaults, and tuning knobs in one place.
How It Works
The rig separates concerns cleanly between two processes:Authorize Telegram
Run
npm start -- --session to scan a QR code and save a GramJS session string. The scraper uses a regular user account — any channel you can read, it can read.Start the trading process
Launch with
--paper or --live plus --entry pointing to your strategy file. The engine starts one Live.background() loop per symbol in the whitelist and opens an HTTP bridge on port 60051.Attach Claude
Run
npm run start:claude to open Claude Code with mcp.servers.json wired in. The stdio MCP server (npx @backtest-kit/mcp) forwards every tool call over HTTP to the trading process.Key Design Principles
Engine Owns Risk
Take-profit, stop-loss, and entry cost are computed engine-side and cannot be overridden by the model or by anything in the feed.
Feed Is Untrusted
Instructions embedded in Telegram posts are data, not commands. The architecture — not the prompt — enforces this boundary.
Full Audit Trail
Every byte the model saw is dumped to disk as markdown before leaving the process. Every position carries a
note with its basis.Crash-Safe Persistence
Open positions, OCO brackets, and command history survive process restarts. Resume exactly where you left off.