AI Crypto Trading with Claude: 5 MCP Workflows That Save Hours

Connecting Claude to your trading account is the easy half. The hard half is knowing what to ask it — because most "AI for crypto" demos collapse the moment you want to do something real. This post is the practical half: five concrete workflows where the VolatiCloud MCP server turns multi-screen dashboard work into one conversation, plus copy-pasteable prompts and the trade-offs nobody warns you about.

Haven't connected MCP yet? Start with the MCP server setup guide — it covers configuration for Claude Code, Claude Desktop, and Cursor, and how the tool architecture works. Come back here once you're connected.

The VolatiCloud MCP server exposes every operation the dashboard exposes — queries, mutations, backtest runs, bot lifecycle commands — to Claude as typed tools. The assistant fetches data, reasons over it, and takes action inside a single conversation. Multi-step workflows that previously required jumping across five dashboard screens become one prompt; Claude chains the calls and pauses at destructive actions until you confirm.

Five Practical AI Trading Workflows

1. Backtesting from a Conversation

With the MCP server active in Claude Code, you can run a full backtesting workflow without leaving your editor:

Me: What strategies do I have?
Claude: [calls listStrategies] You have 4 strategies: ...

Me: Run a backtest on "MACD Crossover v3" for BTC/USDT over the last 90 days
Claude: [calls createBacktest, then polls status] Backtest complete. 
        Win rate: 58%, profit factor: 1.41, max drawdown: 12.3%

Me: That drawdown looks high. What does Monte Carlo say?
Claude: [calls createSimulation] Worst-case drawdown at 95th percentile: 18.7%

Each step is a real API call against your account. The results are not generated — they are your actual backtest data.

2. Multi-Bot Status Reports

Me: Give me a summary of all running bots and their unrealized P&L
Claude: [calls listBots with status filter] 
        3 bots running, 1 in dry-run mode:
        - ETH/USDT Trend Bot: +2.4% unrealized
        - BTC Grid: -0.8% unrealized  
        - SOL Momentum: +5.1% unrealized
        - XRP Test (dry-run): +1.2% simulated

This kind of report previously required opening each bot's detail page. The MCP server fetches it all in one query.

3. Strategy Iteration Loop

The combination of the MCP server and strategy versioning makes iterating on strategies dramatically faster. Inside a single conversation thread you can: fork a strategy, adjust indicator parameters, run a backtest, compare against the parent version's existing backtests, and promote the best variant. The assistant orchestrates the chain — createFork, updateStrategy, createBacktest, bot queries — without you ever opening the dashboard.

This pairs well with hyperparameter optimization: ask Claude to run a hyperopt sweep on the new fork, then summarize the top three parameter sets and their out-of-sample behavior.

4. Pre-Weekend Portfolio Audit

Say you want to review all your active bots before stepping away for the weekend. Without MCP, that means navigating to each bot manually. With Claude:

You: Summarize the 7-day performance of all my active bots. Flag anything 
     with a drawdown above 3% or that hasn't traded in 48 hours.

Claude: [calls listBots, then getBotPerformance for each]

Here's the summary:
- BTC Momentum (live, Binance): +3.2% / 7d, 14 trades, max drawdown 1.1% ✓
- ETH Mean Reversion (live, Kraken): -0.4% / 7d, 8 trades, max drawdown 4.2% ⚠ drawdown
- SOL Breakout (paper, Binance): +6.7% / 7d, 22 trades, max drawdown 2.3% ✓
- ADA Grid (live, Bybit): 0 trades in 72h ⚠ stalled

ETH Mean Reversion and ADA Grid need attention. Want me to pull their recent 
trade logs?

From there you can dig into the problem, run a comparison backtest, or stop the underperforming bot — all from the same conversation.

5. Incident Response

When an alert fires, context-switching to the dashboard costs time. With MCP, you can stay in Claude and investigate:

You: My "ETH Scalper" alert just fired. What's happening?

Claude: [calls getBotTrades, getBotOpenPositions, getBotTradingConfig]

ETH Scalper has 1 open position: long ETH/USDT entered at $3,241 
(now at $3,189, -1.6% unrealized). The last 5 trades show increasing 
slippage — avg 0.18% vs expected 0.05%. The bot has been running 
continuously for 14 days without a restart.

Would you like me to stop the bot, or just monitor and alert you if 
the loss exceeds 2%?

The assistant chains the calls automatically, presents the relevant data, and defers the destructive action to you.

Example Prompts to Get You Started

These are prompts you can paste directly into Claude once the MCP server is connected. Each one triggers one or more real API calls against your account.

Discovery:

• "What strategies do I have, and which ones have been backtested in the last 30 days?"
• "List all my exchanges and show which ones have active bots."

Backtesting:

• "Fork my RSI strategy, bump the RSI period from 14 to 21, and run a backtest on BTC/USDT over the last 6 months."
• "Compare the last two backtest runs on my MACD Crossover strategy. Which had better risk-adjusted returns?"

Bot management:

• "List all bots that have been stopped for more than 24 hours. For each one, show the profit factor from their last backtest."
• "Stop every bot on the Bybit exchange and confirm they've closed all open positions before you do."

Monitoring:

• "Show me the last 10 trades on my BTC Momentum bot. Is the average slippage within the expected 0.1% range?"
• "Which of my live bots has the highest unrealized P&L right now?"

Strategy creation:

• "Create a new strategy called 'EMA Crossover v1' using a 20/50 EMA crossover signal on the 1h timeframe. Deploy it as a dry-run bot on Binance with 100 USDT stake."

What Stays the Same

The MCP server is an additional interface to VolatiCloud — it does not change how your bots run or how strategies are stored. All data lives in your account under normal authorization rules. The AI assistant can only do what your API client's scopes permit.

warning

The AI assistant cannot see your exchange API keys or private credentials. Those are protected by field-level encryption and require the view-secrets scope, which should not be granted to MCP clients.

Bots deployed through the MCP server are identical to those created via the dashboard. They use the same bot runners, the same exchange connections, and produce the same monitoring data you can watch in real time on the dashboard.

Pros and Cons of LLM-Driven Trading Automation

Using an AI assistant to manage trading bots is genuinely useful — and has real limits worth understanding.

Where it excels:

• Multi-step workflows that span multiple resources (list → analyze → act)
• Generating natural-language summaries of complex performance data
• Rapid iteration loops: fork strategy, tweak params, backtest, compare, repeat
• Incident triage when you need to understand a situation before acting

Where to be cautious:

• The assistant cannot make real-time decisions autonomously — each action requires your prompt
• Claude will not execute destructive mutations (stop bot, delete strategy) without explicit confirmation, but you should still review what it's about to do
• The MCP server operates within your API client's permission scopes — narrowly-scoped clients are safer for exploratory sessions
• Exchange API keys and credentials are never visible through MCP (protected by field-level encryption)

The MCP server is a control interface, not an autonomous trading agent. You remain in charge; Claude does the leg-work of fetching, formatting, and sequencing API calls.

Getting Connected

If you haven't set up the MCP server yet, the MCP server setup guide walks through every step: pasting the config into Claude Code / Claude Desktop / Cursor, completing the OAuth flow, and verifying the connection. The MCP overview docs cover the tool architecture and troubleshooting in more depth.

Once connected, the prompts above are a good starting point. For the full list of available GraphQL operations, use search_schema in your assistant or browse the API documentation. And once you're comfortable with conversational control, the AI Chat Assistant inside the dashboard offers the same surface without leaving the browser — same MCP tools, different surface.

Open console.volaticloud.com to grab the MCP config block, paste it into your editor, and run your first conversation-driven backtest in the next ten minutes.