Skip to main content

What Is MCP Trading? A Guide for Futures Traders

Most articles about AI and trading aim at the developer audience. This page is for the futures trader. If you trade MES, ES, NQ, MNQ, or CL on NinjaTrader 8, and you have heard the words "Model Context Protocol" or "MCP" and want a plain-English explanation, start here.

The short version: MCP is the standard that lets an AI assistant (Claude, ChatGPT, Cursor) call typed tools on a remote service. CrossTrade hosts an MCP server at https://app.crosstrade.io/v1/api/mcp that gives AI clients a structured way to inspect your NinjaTrader 8 install, read your CrossTrade trade copier history, review your TradingView webhook signals, write and backtest NinjaScript, and, with the right permissions, place orders.

You do not need to learn the protocol. You need to know what the agent can do, what you should never let it do, and how to ask for the workflows that pay off.

MCP in one paragraph​

Model Context Protocol is a standardized way for an AI model to discover and call tools on an external system. The model asks the server: "what tools are available?" The server returns a typed list. When the model wants to do something, it calls a tool by name with a JSON payload. The server runs the call and returns a result. That is the entire protocol. It replaces the historical pattern of stuffing API documentation into a chat window and hoping the model writes correct HTTP code. In practice, it means an AI assistant can now talk to NinjaTrader 8, your CrossTrade account, your trade copier, and your journal in a way that does not break the next time you upgrade.

Why futures traders should care​

Futures traders live with friction that office workers don't:

  • A trade is rejected and you don't know if it was max contracts, daily loss, or symbol mapping.
  • A TradingView alert fired but no order showed up on the chart.
  • The trade copier filled the leader but skipped a follower.
  • A NinjaScript compile error blocks you for 20 minutes after the open.
  • A funded account is on a knife's edge of trailing drawdown and you're guessing at room.

These are exactly the workflows an AI agent can help with, if the agent can inspect real state instead of guessing. MCP is what gives the agent that access.

The opportunity is not "AI trades for me." The opportunity is "AI tells me what just happened and why."

Read-only AI agents​

The safest and most useful starting point. The agent has mcp:read permission. It can inspect:

  • Accounts and connection state
  • Positions and working orders
  • Executions and fills
  • The matched-trade journal
  • TradingView webhook signal history
  • CrossTrade Add-On activity log
  • Watermarks (daily highs and lows used by trailing drawdown rules)
  • Bars, quotes, and instrument metadata
  • Compiled and on-disk NinjaScript

What you get is a junior analyst who can answer questions like:

  • "Summarize my last 20 trades on Sim101 and group by setup."
  • "Which webhook signals fired today but did not produce an order?"
  • "How close am I to my daily loss limit on APEX1234?"
  • "Did the follower account miss a copier signal in the last hour?"
  • "Show me the three biggest losers this week and what was different about them."

Read-only sessions do not move anything in your account. They cannot. The token does not allow it.

Trade-enabled AI agents​

When you upgrade the agent's scope to mcp:trade, it can:

  • Place, modify, and cancel orders
  • Flatten positions
  • Compile and write NinjaScript
  • Run Strategy Analyzer backtests
  • Sweep parameters
  • Deploy and stop strategies
  • Emit alerts to your in-app, email, Discord, or Telegram

Trade-enabled is more powerful, and proportionally more dangerous. The first principle: scope is the floor, not the ceiling. Even with mcp:trade, your prompt should force the agent to:

  1. Read state first (accounts, positions, orders, watermarks).
  2. Restate the proposed action with account, instrument, side, and quantity.
  3. Wait for your explicit confirmation.
  4. After the action, verify the result.

This is not over-cautious. This is what experienced traders do manually. The agent should be no different.

Why NinjaTrader 8 needs an add-on bridge​

NT8 is a Windows desktop application. It is not a cloud broker with a REST API. To reach NT8 from anywhere off your machine, you need a bridge that runs inside NT8. That is the CrossTrade Add-On. It connects out to CrossTrade's hosted services. Without the add-on, no external API, no MCP, no webhooks. With the add-on, NT8 becomes accessible the same way a cloud service would be.

This is why CrossTrade exists. It is also why MCP for NinjaTrader needs more than a generic local MCP server: somebody has to talk to NT8, and that piece runs on your machine.

Examples of MCP trading workflows​

Pre-trade risk brief​

Before I trade APEX1234 this morning, give me a risk brief: accounts, connection state, open positions, working orders, account summary, watermarks. Tell me how much room I have under the $1,500 trailing drawdown. Do not place anything.

The agent calls ListAccounts, GetConnections, ListPositions, ListOrders, GetAccountSummary, GetWatermarks. You get a one-page brief in 10 seconds.

Journal review​

Pull the last 20 trades on APEX1234. Group by instrument and setup. Tell me win rate, average win, average loss, the three biggest losers, and any pattern in their entry times. Redact the account number in the summary.

The agent calls GetJournalTrades and produces an analysis. No write permissions needed.

Webhook diagnostics​

Pull TradingView webhook signals for Sim101 in the last 24 hours. For each, tell me the intended action and whether it produced an order. Highlight any that did not fill and the likely reason.

The agent calls GetSignalHistory, ListOrders, ListExecutions, and GetActivityLog. You see the full path from alert to no-fill.

Trade copier diagnostics​

For APEX1234 and APEX1235 between 14:30 and 14:35 today, pull the add-on activity log. Reconstruct the timeline. Tell me why the follower missed the leader's fill.

The agent reads the activity log and timestamps. You learn whether the cause was max contracts, account state, or a symbol mapping issue.

NinjaScript compile loop​

Write a 9/21 EMA cross strategy for MES with an ATR-based trailing stop. Compile in memory. If compile fails, fix and recompile. Do not write the file unless I confirm.

The agent calls GetNinjaScriptHelp, LookupNinjaScriptSymbol, CompileNinjaScript(in_memory: true), repairs as needed, and pauses for confirmation.

Risk and funded account constraints​

Funded futures rewards discipline. AI does not change that. A few specifics:

  • Many prop firms restrict or prohibit automation. Apex Trader Funding's current prohibited activities list explicitly bars automation or algorithm usage. Always check the official firm page before assuming an agent can act for you.
  • Trailing drawdown is brutal. Your high watermark moves up but never down. AI cannot save you from a bad fill that breaches it. The best use of AI on funded accounts is read-only inspection and pre-trade checklists.
  • Daily loss limits are real. The agent should refuse to place new trades when room is exhausted.
  • News windows. TakeProfitTrader PRO and similar accounts have firm-specific news-window prohibitions. Bake the rule into the system prompt.

See AI for Funded Futures Traders and the individual firm pages for specifics.

How CrossTrade MCP fits​

CrossTrade owns three things in this loop:

  1. The hosted MCP server. One URL, OAuth, dynamic client registration, every tool typed.
  2. The CrossTrade Add-On inside NT8. The bridge that gives the server real access to your install.
  3. The journal, copier, webhook, and add-on activity history. The data an AI agent needs to actually explain what happened.

A generic "trading MCP server" can place orders. CrossTrade MCP can place orders, and read the months of history needed to know whether placing an order is a good idea today.

Glossary​

  • MCP: Model Context Protocol. A standard for AI clients to call tools on a server.
  • Agent: A program built on top of an AI model that calls tools, observes results, and decides next steps.
  • Scope: A token's permission level. CrossTrade MCP uses mcp:read and mcp:trade.
  • State check: Reading account, position, and order state before any write action.
  • Approval gate: A point in the prompt where the agent must wait for your explicit confirmation.
  • NinjaScript: NinjaTrader 8's C# scripting language for indicators and strategies.
  • Strategy Analyzer: NT8's built-in backtesting engine.
  • Watermark: The highest or lowest account balance reached today, used to compute drawdown room.

FAQ​

Can AI trade my NinjaTrader account?​

Technically yes, with mcp:trade scope and explicit confirmation. Practically, the safe answer for most traders is "no, use it read-only first." Build trust with inspection workflows before you grant write permissions.

Is MCP the same as a trading bot?​

No. A trading bot runs continuously on a fixed strategy. An AI agent is a program that calls tools and decides next steps. The agent can use a tool that places an order, but only when its prompt and its tokens allow it.

Does this work for funded accounts?​

It depends on the firm. For funded accounts, the default should be read-only. If you do enable trading, force the agent to verify daily loss room, max contracts, and news-window rules before every action. Some firms explicitly prohibit automation.

What does "Elite subscription" mean?​

CrossTrade's MCP transport is only available on the Elite tier. REST and WebSocket remain on Pro. The Elite gate is on the protocol, not the data.

What is the add-on version requirement?​

CrossTrade Add-On v1.13.0 or newer is required for MCP. Older add-ons either return reduced capability or no MCP tools at all.

What if the agent gets something wrong?​

Two layers protect you. First, scope: a mcp:read token cannot place orders no matter what the agent decides. Second, approval gates in the prompt: a good prompt requires the agent to restate and wait for your confirmation before any write. Use both.

Next steps​