mcp-code-exec

MCP Code Execution Pattern - Implementation Guide

Your task is to convert an MCP server into the Code Execution Pattern described in Anthropic's blog post. This pattern enables progressive disclosure of tools, reducing token usage by up to 98% compared to loading all tools upfront.

---

Why This Pattern?

Problem with Traditional MCP:

• Loading all tools upfront consumes 150,000+ tokens
• Intermediate tool results flow through model context repeatedly
• Agents slow down with many connected tools

Solution - Code Execution Pattern:

• Tools organized as importable Python modules (filesystem-based)
• Agents load only what they need (progressive disclosure)
• Data processing happens in code before returning to model
• Result: 98.7% token reduction (150K → 2K tokens)

---

Step-by-Step Implementation

Step 1: Create Folder Structure

Create a dedicated folder for the MCP server in ./servers/ directory:

./servers/[server_name]/
├── server.py           # Server singleton & connection management (with discovery in __main__)
├── [tool_1].py         # Individual tool file (with test in __main__)
├── [tool_2].py         # Individual tool file (with test in __main__)
├── [tool_3].py         # Individual tool file (with test in __main__)
└── __init__.py         # Package exports

Example:

./servers/notion/
├── server.py
├── notion_search.py
├── notion_fetch.py
├── notion_create_pages.py
└── __init__.py

Create the folder:

mkdir -p ./servers/[server_name]

---

Step 2: Create the Server Module

The server.py module manages the MCP server connection as a singleton. Include tool discovery in the __main__ block to list available tools.

Note: This pattern works with any MCP server type. The template below shows MCPServerStdio (most common), but you can use:

• MCPServerStdio - Local scripts, npm packages, or OAuth servers with mcp-remote
• MCPServerSse - Server-Sent Events servers
• MCPServerStreamableHttp - HTTP streaming servers
• HostedMCPTool - Publicly accessible servers (no need for singleton pattern)