Alpaca MCP Split Server

Two MCP servers for Alpaca trading — split by purpose to work within Claude's tool limits.

The Problem

Claude Desktop (and claude.ai) has a hard limit on how many MCP tools it will load per conversation. When a single MCP server exposes 40+ tools, Claude silently drops many of them. You end up with a powerful server that only partially works — often missing the exact tools you need most (like order placement or options chains).

There's no error message. Tools just don't appear. You might have get_account_info, get_positions, place_stock_order, get_option_chain, and 35 other tools in your server — but Claude only loads ~12-15 of them, seemingly at random.

The Solution

Split one large server into two smaller ones organized by purpose:

Claude loads tools from all connected MCP servers into one flat pool. Two servers with 30 and 9 tools each both stay under the limit, giving you access to all 39 tools in every conversation.

Both servers connect to the same Alpaca account using the same API keys. There's no state isolation — they're just two entry points to the same brokerage account.

What Was Removed

All crypto trading tools (9 tools) were removed from the original monolith to further reduce tool count. If you need crypto, you can add them back to the data server from the original source.

Setup

1. Clone and install dependencies

git clone https://github.com/Schmoll86/alpaca-mcp-split.git
cd alpaca-mcp-split

# Create a shared virtual environment
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Configure API keys

Copy the example env file into both server directories:

cp alpaca-trading-data/.env.example alpaca-trading-data/.env
cp alpaca-trading-orders/.env.example alpaca-trading-orders/.env

Edit both .env files with your Alpaca API credentials. You can get these from Alpaca Dashboard.

3. Configure Claude Desktop

Add both servers to your claude_desktop_config.json:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "alpaca-trading-data": {
      "command": "/path/to/alpaca-mcp-split/venv/bin/python",
      "args": [
        "/path/to/alpaca-mcp-split/alpaca-trading-data/alpaca_trading_data.py"
      ],
      "env": {
        "PYTHONUNBUFFERED": "1"
      }
    },
    "alpaca-trading-orders": {
      "command": "/path/to/alpaca-mcp-split/venv/bin/python",
      "args": [
        "/path/to/alpaca-mcp-split/alpaca-trading-orders/alpaca_trading_orders.py"
      ],
      "env": {
        "PYTHONUNBUFFERED": "1"
      }
    }
  }
}

Replace /path/to/alpaca-mcp-split/ with your actual clone path.

4. Restart Claude Desktop

Quit and relaunch. Both servers will start automatically. In a new conversation, you should have access to all 39 tools.

Tool Inventory

Data Server (30 tools)

Account & Positions: get_account_info, get_positions, get_open_position

Stock Data: get_stock_quote, get_stock_bars, get_stock_trades, get_stock_latest_trade, get_stock_latest_bar, get_stock_snapshot, get_stock_quotes

Options Data: get_option_contracts, get_option_latest_quote, get_option_snapshot, get_option_chain

News: get_news, get_latest_news

Market: get_market_clock, get_market_calendar

Watchlists: create_watchlist, get_watchlists, update_watchlist, get_watchlist_by_id, add_asset_to_watchlist_by_id, remove_asset_from_watchlist_by_id, delete_watchlist_by_id

Other: get_asset_info, get_all_assets, get_portfolio_history, get_corporate_announcements, get_session_state

Orders Server (9 tools)

Stock Orders: place_stock_order, get_orders, cancel_all_orders, cancel_order_by_id

Position Management: close_position, close_all_positions, exercise_options_position

Options Orders: place_option_market_order

Session: record_decision

How It Works

┌─────────────────────────────────────────────┐
│              Claude Desktop                  │
│                                              │
│  Tools from ALL connected MCPs appear as     │
│  one flat pool. Claude picks the right       │
│  tool regardless of which server owns it.    │
└──────────┬──────────────┬────────────────────┘
           │              │
    ┌──────▼──────┐ ┌─────▼───────┐
    │  Data MCP   │ │ Orders MCP  │
    │  (30 tools) │ │  (9 tools)  │
    └──────┬──────┘ └─────┬───────┘
           │              │
           └──────┬───────┘
                  │
         ┌────────▼────────┐
         │  Alpaca API     │
         │  (Same account, │
         │   same keys)    │
         └─────────────────┘

Adapting This Pattern

This same split strategy works for any MCP server hitting Claude's tool limit:

1. Count your tools — if you have 30+, you're probably getting silently truncated
2. Split by read vs. write — safest separation, no functional dependencies
3. Share the venv — symlink or use the same virtual environment for both
4. Same .env — both servers use identical credentials

License

MIT