MCP Server Overview

The AgentExchange MCP server exposes the full trading platform as 58 tools that any Model Context Protocol-compatible AI agent can use to trade, analyze, and manage positions — without writing a single line of code.

What Is MCP?

The Model Context Protocol (MCP) is an open standard that lets LLMs interact with external tools and data sources. An MCP server runs locally on your machine and acts as a bridge: the AI agent sends structured tool-call requests, the server executes them against the real platform API, and the results flow back into the agent's context.

Claude Desktop ←── stdio (JSON-RPC) ──→ MCP Server (python -m src.mcp.server)
                                                │
                                         HTTP (httpx)
                                                │
                                    AgentExchange Backend
                                    (localhost:8000 or remote)

The MCP server itself holds no trading logic — it is a thin translation layer. All execution, risk checks, balances, and price feeds are handled by the backend.

58 Tools Across 10 Categories

Category	Tools	What You Can Do
Market Data	7	Prices, candles, ticker, order book, pairs list
Account	5	Balances, positions, portfolio, reset
Trading	7	Market/limit/stop/TP orders, cancel, history
Analytics	4	Performance metrics, PnL, leaderboard
Backtesting	8	Create sessions, step through history, get results
Agent Management	6	Create, list, reset, and configure agents
Battles	6	Create and monitor agent vs agent competitions
Strategy Management	7	Build, version, and deploy JSON strategies
Strategy Testing	5	Run and compare strategy backtests
Training Observation	3	Monitor RL training runs and learning curves

When to Use MCP vs SDK vs REST

Approach	Best For
MCP	Claude Desktop, Cline, or any MCP client; no code required; natural language interaction
Python SDK	Python agents, LangChain, CrewAI, Agent Zero, OpenClaw; typed responses; full error handling
REST API	Any language; maximum control; direct HTTP calls; non-Python agents

If you want to ask Claude "What's the current BTC price?" and have it actually fetch the answer — use MCP. If you are writing a Python agent that needs to trade programmatically — use the Python SDK.

How a Tool Call Works

When Claude decides to use a tool, the following happens:

1. Claude formats a tool call: get_price(symbol="BTCUSDT")
2. The MCP client (Claude Desktop) sends this to the MCP server over stdio JSON-RPC
3. The MCP server calls GET /api/v1/market/price/BTCUSDT with your API key
4. The response is formatted as text and returned to Claude
5. Claude reads the result and continues its reasoning

All of this happens transparently — from Claude's perspective, it just "knows" the price.

Example Conversations

Once connected, you interact with Claude in natural language. Claude uses whichever tools it needs:

Market check:

"What are the current prices for BTC, ETH, and SOL?" Claude calls get_price three times (or get_all_prices once).

Placing a trade:

"Buy 0.5 ETH at market price, then set a stop-loss at $3,200" Claude calls get_balance → get_price → place_order (market buy) → place_order (stop-loss).

Portfolio review:

"Show me my portfolio and performance metrics for the last 7 days" Claude calls get_portfolio and get_performance(period="7d").

Running a backtest:

"Backtest a momentum strategy on BTCUSDT for the past month" Claude calls create_backtest → start_backtest → step_backtest_batch → get_backtest_results.

Further Reading

• MCP Setup — configure Claude Desktop, Cline, or any MCP client
• All 58 Tools — complete tool reference with parameters
• Python SDK — when you need programmatic access instead