An MCP trading server is a remote service that gives an AI app such as Claude or ChatGPT a set of trading tools it can call: market data, strategy validation, backtests and paper bots. MCP stands for Model Context Protocol, an open protocol that lets an AI app call tools on a server someone else runs. The model writes the strategy and asks for the test. The server does the work and sends back the result.
This guide explains the protocol in plain words, what a trading server should and should not expose, how to connect the OptAlgo connector, and what to check before you trust any trading MCP server with your account.
What the Model Context Protocol is
A language model on its own can only produce text. It cannot load four years of Bitcoin candles, run a backtest or open a position. The Model Context Protocol fills that gap. A server publishes a list of tools, each with a name, a description and the inputs it accepts. The AI app reads that list, and when your request needs one of the tools, the model calls it and reads the answer.
The protocol is open, so one server works in several apps. You add a server by its URL, sign in, and the tools appear in your chat. Nothing is installed on the exchange side, and the model does not need to know how the server works inside. It only needs the tool list.
The important consequence: the model can only do what the server offers. If a server has no tool that places a live order, no prompt can make the model place one. The limits live on the server, not in the conversation.
What a trading MCP server exposes
Servers differ, but a useful trading server covers the steps between an idea and a bot that runs. The OptAlgo MCP connector is one example. Its tools let your AI:
- List markets and history, and check live prices.
- Check your plan limits and read the strategy SDK, so the code it writes fits the platform.
- Validate a strategy for free before any quota is spent.
- Run backtests and read the honesty grade, the overfit verdict and the equity chart.
- Run research such as a walk-forward, which judges a strategy on data the optimiser never saw.
- Create paper bots, open and close paper positions, and list trades.
Validation deserves a note. It queues nothing, loads no market data and costs nothing. It runs the code once on synthetic bars and lists every problem at once, including lookahead and non-deterministic code, each with its fix. A model that validates first wastes far fewer runs.
Why paper-only by design matters
A chat model can misread a request, repeat a call or act on a sentence you did not mean as an order. With real money that is a real loss. With paper trading it is a line in a log.
The OptAlgo connector is paper-only by design. It cannot place a live order, flip a bot to live or touch a live bot. This is not a rule written in a prompt that the model is asked to follow. The tools for live trading are not there. Going live is your decision, made in the OptAlgo app, on a Plus or Pro plan.
When you compare servers, this is the first thing to look for. A server that can place live orders from a chat puts a great deal of trust in a text conversation.
OAuth instead of pasted keys
Some setups ask you to paste an API key into a chat or a config file. A key in a chat is in the conversation history, and anything that reads the history can read the key.
A remote MCP server can use OAuth instead. You add the URL, the app sends you to the service's own sign-in page, you approve, and the app receives access without ever seeing a password or a key. The OptAlgo connector works this way: you never paste an API key into a chat, and you can disconnect it in one click under AI agents in the app. Access stops immediately.
How to add the OptAlgo connector
You need a free OptAlgo account: sign up here. No card and no exchange account are needed, because everything the connector does runs on paper.
In Claude.ai (web, desktop or mobile) and in ChatGPT, add https://mcp.optalgo.com/mcp as a custom connector, sign in to OptAlgo and approve. In Claude Code, run this once in your terminal:
claude mcp add --transport http optalgo https://mcp.optalgo.com/mcp
Then type /mcp inside Claude Code, pick optalgo, sign in and approve. The step-by-step setup for each app is in the connector docs. Codex and other terminal agents use a different path, a device login and the REST API, which the Claude Code guide covers.
The connector is free on every plan, including Free. Only running a backtest uses quota. Validation, market data and paper trading do not.
What a session looks like
You describe an idea in one sentence. For example: “Go long BTC on 4h when the 20 EMA crosses above the 80 EMA, with a 3 ATR trailing stop. Test it from 2022 to the end of 2025.”
- The model reads the SDK and writes the strategy as plain Python.
- It validates the code for free and fixes every finding.
- It runs the backtest on real exchange data.
- It reports the honesty grade and the overfit verdict first, then the numbers.
- If the result is worth testing further, it asks for a walk-forward research run.
- If you want, it deploys the strategy as a paper bot that runs at every bar close.
Here is a real result from that exact prompt, run on OptAlgo on 2026-11-22 with Binance data. The strategy was an EMA cross trend strategy (long when EMA 20 crosses above EMA 80, 3 ATR trailing stop) on the BTCUSDT perpetual, 4h bars, 2022-01-01 to 2026-01-01.
- Honesty grade: A.
- Return after fees and funding: +70.8% (CAGR 14.3%).
- Max drawdown: -15.0%. Sharpe: 0.85.
- 55 trades, win rate 40%, profit factor 1.79.
Read the grade before the return. An honesty grade of A means every fill settled on real data with fees, funding and exchange rules. It says the backtest is honest about fills. It does not say the strategy has an edge.
This is one backtest, and one backtest is in-sample. A walk-forward research run of the same idea, with 6 rolling folds from 2021-01-01 to 2026-01-01, came back with the verdict likely overfit: the out-of-sample return was +23.4% with an out-of-sample Sharpe of 0.49, and it failed the deflated Sharpe check. So a grade A run with +70.8% is a starting point for more testing, never proof. A good trading server makes that second step easy, and a good session includes it. The overfitting guide explains how to read these verdicts.
What to check before trusting a trading MCP server
Anyone can publish an MCP server. Before you connect one to an account that matters, answer these questions.
- Who runs it? Look for a named company, a contact address and documentation. An anonymous server that asks for exchange keys is a risk you do not need.
- What can it do with your account? Read the tool list. Check whether any tool places live orders, changes live bots or moves funds. Paper-only is the safest starting point.
- How does it authenticate? OAuth with a sign-in page you recognise is better than a key pasted into a chat or a config file.
- Does it ever need withdrawal permission? It should not. A bot trades, it does not withdraw. See the API key permissions guide.
- Are backtests graded? A server that returns only a return figure gives a model nothing to doubt. Look for a realism grade, fees and funding in the result, and a check for lookahead.
- Does it count your tries? Fifty variants on the same data produce one lucky winner. A server should track this and offer walk-forward testing.
- Can you disconnect? You should be able to revoke access in one step, and it should take effect at once.
- Where do the limits live? Limits written in a prompt can be talked around. Limits enforced by the server cannot.
For OptAlgo the answers are public. It is operated by FreedX Technologies OÜ, an Estonian company. The connector is paper-only and uses OAuth. Every backtest comes with an A to D honesty grade and an overfit verdict based on a deflated Sharpe ratio. The security page lists what OptAlgo can and cannot do with your account.
A plain warning
Trading is risky, and a model that writes clean code quickly does not change that. An MCP trading server makes testing an idea faster and more honest. It does not make the idea profitable. Use the connector to find out which ideas fail, keep the ones that survive out-of-sample tests, and run them on paper before you risk money.
Frequently asked questions
What is an MCP trading server?
It is a remote server that gives an AI app such as Claude or ChatGPT trading tools through the Model Context Protocol. Typical tools are market data, strategy validation, backtesting and paper bots. The model calls the tools and the server does the work.
Can Claude trade crypto through an MCP server?
It depends on the server. The OptAlgo connector is paper-only: Claude can write, validate and backtest strategies and run paper bots, and it can never place a live order. Live trading happens only in the OptAlgo app, after you confirm.
Do I need to paste an API key to use the OptAlgo MCP connector?
No. The connector uses OAuth. You add the URL, sign in to OptAlgo and approve. No API key is pasted into a chat, and you can disconnect any time under AI agents in the app.
Is the OptAlgo MCP connector free?
Yes, on every plan including Free. Only running a backtest uses quota. Validation, market data and paper trading do not.
Does a good backtest from an MCP server mean the strategy works?
No. One backtest is in-sample. A high honesty grade means the fills are realistic, not that the strategy has an edge. Run a walk-forward and a paper bot before you decide anything.