Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/backtest-kit/ai-trading-mcp/llms.txt

Use this file to discover all available pages before exploring further.

AI Trading MCP is a Model Context Protocol server that gives Claude (or any MCP client) a minimal, guarded vocabulary for crypto trading: get_status, open_position, and close_position. The LLM is the only signal source — a GramJS scraper pipes a live Telegram channel, including chart screenshots, directly into the agent’s context on every get_status call. The trading engine owns all risk parameters: take-profit, stop-loss, position sizing, and the symbol whitelist are computed and enforced engine-side, not via prompt engineering.

Quickstart

Install, authorize Telegram, and run your first paper trade in minutes

Architecture

Understand the two-process design and how the MCP bridge works

MCP Tools

Full reference for get_status, open_position, and close_position

Live Broker

How the battle-hardened Binance spot adapter handles real orders

How It Works

The rig runs as two separate processes connected over HTTP. Claude speaks to a stateless stdio MCP server; that server forwards every call to the trading process, which manages exchange connections, position state, and the Telegram feed.
1

Authorize Telegram

Run npm start -- --session to scan a QR code and save your GramJS session. The scraper reads the channel as a real user account.
2

Start paper trading

Run npm start -- --paper --entry ./content/manual.strategy/manual.strategy.ts --ui to launch 13 live-data symbol loops and the dashboard on port 60050.
3

Attach Claude

Run npm run start:claude to open Claude Code with the MCP server wired in via mcp.servers.json. The agent can now see your portfolio and the Telegram feed.
4

Go autonomous

Use Claude’s /loop 15m command to run the trading prompt on a timer — the rig becomes an unattended news-trader with a full audit trail of every decision.

Key Design Principles

Engine-Owned Risk

TP/SL and entry cost are computed by the backtest-kit engine and cannot be overridden by the model. The agent’s vocabulary has no size or leverage parameters.

Telegram as Sensory Organ

The last 15 channel posts — text and chart screenshots — are included in every get_status response as native MCP image blocks.

Full Audit Trail

Every get_status response is dumped to disk as markdown before it leaves the process, with images referenced relatively for the built-in dump viewer.

Adversarial Feed by Design

The demo channel is a real scam signals channel. The architecture — not the prompt — bounds the blast radius of any manipulated post.
This project requires a Telegram API key pair from my.telegram.org and a GramJS user session (not a bot token). For live trading, Binance API credentials are also needed. Paper mode runs with no exchange keys.

Build docs developers (and LLMs) love