MCP

CrossTrade MCP Is Live: The Hosted AI Trading Server for NinjaTrader 8

CrossTrade MCP is now live at https://app.crosstrade.io/v1/api/mcp. A hosted Model Context Protocol server for NinjaTrader 8, with OAuth, scoped permissions, and the CrossTrade Add-On bridging to your real install.

CrossTrade Team

• 3 min read
Share
crosstrade-mcp-is-live-ai-hosted-trading-server-for-ninjatrader-8-black-background-purple-highlight

For the past year, AI assistants have been able to chat about trading. Until now, they have not been able to do much trading. The wall has been ergonomic. NinjaTrader 8 lives on a Windows desktop. AI clients live in a browser or a CLI. Without a bridge, the conversation could only ever be about trading, not the trading itself.

That bridge is what CrossTrade has spent years building for webhooks, REST, and our trade copier. Today we are turning it into a Model Context Protocol server.

CrossTrade MCP is live at https://app.crosstrade.io/v1/api/mcp.

What shipped

The launch covers:

  • Hosted MCP endpoint. One URL. Streamable HTTP transport. Compatible with any MCP client that follows the standard discovery flow.
  • OAuth 2.1 with PKCE and dynamic client registration per RFC 7591. Most clients only need the URL; the rest is automatic.
  • Two scopes: mcp:read for inspection-only workflows, mcp:trade for writes.
  • Elite-tier gate. MCP is part of the Elite product. REST and WebSocket remain on Pro.
  • CrossTrade Add-On v1.13.0 or higher required inside NT8. Older add-ons return a reduced or empty tool surface.

Verified at launch:

  • NT8 Strategy Analyzer parity. Single backtests through RunStrategyBacktest are bit-identical to NT8's UI for the documented reference parameters.
  • The full NinjaScript compile loop: help, search, lookup, in-memory compile, file write.
  • The matched-trade journal, TradingView webhook signal history, and CrossTrade Add-On activity log exposed as read-only tools.
  • Deployment registry: deploy, verify, stop.

Why NinjaTrader needed a real agent bridge

An AI agent is not a chatbot. It is a program that calls typed tools, reads results, and decides next steps. Without typed tools, the agent guesses. With typed tools that talk to your real install, the agent converges. The reason most AI NinjaScript work has been frustrating is that the model has never been able to look at your actual NT8. It hallucinates indicators that do not exist and method overloads that were removed two versions ago.

CrossTrade MCP gives the agent four things it has never had: a help surface, a symbol search, an in-memory compile, and a backtest engine that matches the desktop UI. The compile loop now lives inside the conversation. Compile errors that used to consume a whole morning resolve in three or four iterations.

What a trader can ask now

Read-only is the place to start. Even with no write scope, a CrossTrade MCP agent can:

  • Give you a pre-session risk brief: accounts, positions, working orders, summary, watermarks, remaining drawdown room.
  • Pull your last 20 closed trades, group by setup, identify revenge patterns and biggest losers.
  • Audit your TradingView webhook signals and explain which fired but did not fill.
  • Reconstruct a trade copier incident timeline across leader and follower accounts and surface the verbatim rejection reason.

None of those is a trade. All of them are the work funded traders do manually in a spreadsheet at the end of a session. The agent does them in seconds and without the manual transcription errors.

What a developer can build now

With mcp:trade, the agent can compile NinjaScript in memory, repair failures with LookupNinjaScriptSymbol, write the file after a clean compile, run a Strategy Analyzer backtest, sweep parameters, and deploy with gates. The deployment registry tracks what the agent has launched; GetDeployedStrategyState reports live state.

For developer setup, see the MCP setup docs. For trader-facing prompt examples, see MCP Trading with AI Agents.

Why safety comes first

Trading is unforgiving in a way that office software is not. CrossTrade MCP enforces safety in three layers:

  1. Scope. The OAuth scope makes mcp:read physically incapable of placing an order. Even if the prompt asks for it, the server returns insufficient_scope.
  2. State checks. The recommended prompt pattern requires the agent to read accounts, positions, orders, and watermarks before any write.
  3. Confirmation gates. The recommended prompt requires the agent to restate any write action and wait for an explicit user confirmation.

For funded futures accounts, the firm's rules are binding. Apex Trader Funding's current prohibited activities list bars automation. Topstep permits automated strategies in the Combine with caveats. The agent's job is to encode the rule for you and refuse when room is thin.

How this differs from a generic trading bot

A trading bot runs a fixed strategy continuously. An AI agent is a program that calls tools, observes results, and decides next steps. A bot does not get smarter when you talk to it. An agent connected to MCP gets dramatically more useful when you tell it what to inspect.

CrossTrade MCP is built around the operational reality of NT8: the desktop platform, the add-on bridge, the journal, the webhook history, the copier. A generic broker MCP can place an order. CrossTrade MCP can place an order, and read months of history to know whether placing one is a good idea today.

Start here

If you have wanted Claude or ChatGPT to actually understand what is on your screen in NT8, this is the moment. Start read-only. Build trust on inspection workflows. Upgrade scope only for the specific workflows that need it.