MCP for Trading: control your trading software with Claude and ChatGPT

What MCP is, in one paragraph

Short answer: the Model Context Protocol (MCP) is an open standard that lets an AI assistant call functions in another program. The program runs a small MCP server that publishes a list of tools, each with a name, a description and typed arguments. The assistant reads that list, decides which tool answers your request, calls it and turns the result into an answer. For a trading platform, the tools are things like “list strategies”, “get performance” or “stop strategy”.

The important part for traders: the assistant does not get a screen or a mouse, and it does not guess. It can only call the tools the server publishes, with the arguments the server accepts, and the server decides what each tool is allowed to change. That makes MCP very different from “giving an AI remote access to your VPS”.

Why traders want an AI agent connected to their platform

Algorithmic and arbitrage traders run their software around the clock, usually on a VPS close to the broker or exchange. Most of the day-to-day work is checking, not trading: is every session connected, is every strategy running, what does the log say after a disconnect, how did a strategy do this week, what was the slippage. Each check means a remote desktop session, several windows and some scrolling.

• Status in one question. “Which sessions are disconnected and which strategies are stopped?” replaces a tour of four windows.
• From anywhere. Claude and ChatGPT have mobile apps with voice, so you can check or stop a strategy from your phone without a remote desktop client.
• Analysis, not just data. The assistant reads closed orders, performance and logs and can summarise them: “compare the slippage of my two LockCL2 strategies this week”.
• Routine changes with checks. “Clone strategy 18 and set MaxOrdersTotal to 5” is validated by the server before anything is saved.
• News handling. Refresh the economic calendar, switch events on or off, or stop strategies before a release.

How the connection works: local and VPS

There are two typical layouts. Which one you use decides whether you need a certificate and firewall rules.

The VPS layout is the most common for SharpTrader users: the platform keeps its low-latency location, and the assistant runs where you are. Your phone then talks to the assistant, and the assistant’s desktop app talks to the VPS.

What an AI agent should be allowed to do

Language models are good at reading and summarising and sometimes wrong about what you meant. A sensible design gives an agent wide read access and narrow, explicit write access. In SharpTrader the tools are split into four permission groups:

Just as important is what is not available to any agent: sending a new order (including manual orders), deleting or renaming strategies, adding or removing instruments, session properties and symbol mapping, backtests and optimisation, and anything about paths, files, DLLs or network addresses. Those stay in the program’s own windows.

Security checklist

An MCP server on a trading VPS is a door into your trading. These are the protections to look for, in SharpTrader or in any platform you connect:

Step by step: the server side

In SharpTrader open Settings › AI agents (MCP). For the local layout, tick “Allow AI agents to connect”, keep “This computer only” and press Apply changes. For the VPS layout:

If your VPS provider has its own firewall in its control panel, open TCP port 47350 there too; SharpTrader can only change Windows Firewall.

Connecting Claude

Claude Desktop is the easiest client because SharpTrader ships a ready extension file, SharpTrader.mcpb. It uses Claude Desktop’s own runtime, so nothing else has to be installed.

Given a certificate file, the extension connects only to a server that presents that certificate, so the token never goes anywhere else. If SharpTrader starts after Claude, the extension retries every 15 seconds and loads the tools once it connects. Claude Code, VS Code and Cursor users can instead press the matching Copy … config button in SharpTrader.

Connecting ChatGPT

The ChatGPT desktop app starts local MCP servers as programs, so it reaches SharpTrader through the small mcp-remote bridge. Install Node.js on your computer first (the Windows installer, default options).

Then open ChatGPT › Settings › Plugins, add an MCP server and fill in:

Save, start a new chat and ask “show me my SharpTrader strategies”. If ChatGPT asks whether to trust the folder, confirm. For OpenAI’s hosted agents (Responses API, Agents SDK, Agent Builder) use Copy OpenAI config in SharpTrader instead; they need a Let’s Encrypt certificate and a public address.

Useful requests to try

Good habits: name strategies by ID, ask the agent to read the state again after an action (connecting, closing positions and order updates continue in the background), and keep write requests short and explicit.

Troubleshooting

Limits and honest caveats

• The AI can misunderstand. Keep write permissions off until you need them, and check the journal after actions.
• It is not a trading advisor. An assistant summarising your performance is not a forecast, and it cannot make a strategy profitable.
• Actions are asynchronous. “Started” means the command was accepted; the agent should read the state again to confirm.
• Your assistant’s account is a key. Anyone who can use your Claude or ChatGPT account with the connector installed can use the permissions you granted. Protect it with two-factor authentication.
• Some things stay manual by design: new orders, deleting strategies, sessions and symbol mapping, backtests and optimisation.

FAQ

Trading involves risk. This guide is educational and not financial advice. AI assistants can make mistakes; always review what you allow an agent to do.