🚀 LangGraph CodeAct MCP Example

A powerful, cost-efficient virtual assistant built with LangGraph, LangChain, and Model Context Protocol (MCP) that demonstrates how to efficiently process large tasks through code generation.

Video Overview

Key Benefits

• LLM Cost Reduction - CodeAct approach generates and executes code instead of calling tools one by one, dramatically reducing token usage
• Reliable Task Execution - Handles complex tasks requiring hundreds of actions with minimal supervision
• Context Persistence - Maintains variable state between interactions, enabling multi-step processes without regenerating data
• Safe Code Execution - Uses PyodideSandbox for secure Python execution in a controlled environment

Project Setup

Clone the repository:

git clone https://github.com/n-sviridenko/langgraph-codeact-mcp-example
cd langgraph-codeact-mcp-example

Install dependencies:

poetry install

Create a .env file and update with your API keys:

cp .env.example .env

Run the application:

# Start MCP server
poetry run python -m dotenv run python server.py

# Start LangGraph server
poetry run python -m dotenv run langgraph dev

Architecture Overview

The system consists of two main components:

1. Composite MCP Server - Aggregates multiple MCP services through a unified SSE endpoint
2. Virtual Assistant Graph - Implements both CodeAct and ReAct approaches with dynamic selection

CodeAct vs ReAct Approach

• ReAct Approach - Traditional agent pattern that calls tools one by one

  • Higher cost due to growing context window
  • Can be unreliable for large tasks due to potential early termination

• CodeAct Approach - Agent generates Python code that executes tools programmatically

  • Dramatically reduces token usage by replacing multiple tool calls with code
  • More deterministic behavior with loops and error handling
  • Maintains variable state between executions
  • Uses Claude for reflection and code validation

System Components

• MCP Server - Composite FastMCP server that mounts multiple services under a single endpoint

  • Exposes tools through Server-Sent Events (SSE) protocol
  • Enables secure access from PyodideSandbox environment

• Code Execution Environment - PyodideSandbox WebAssembly-based Python runtime

  • Maintains variable state using thread-based sessions
  • Auto-generates Python function bindings for all MCP tools

• Virtual Assistant Graph

  • Dynamically selects between CodeAct and ReAct based on task
  • Optional tool filtering for specific tasks
  • Pre-configured with Notion API and Google Maps services

Extending the System

Add new MCP services to the services dictionary in setup_mcp_proxy_servers function:

services = {
    # ... existing services
    "new_service": {
        "package": "@package/new-service-mcp",
        "args": [],
        "env": {
            "NEW_SERVICE_API_KEY": os.getenv("NEW_SERVICE_API_KEY", "")
        }
    }
}

Limitations

• Container Persistence - Session data is lost when containers are restarted in cloud environments
• Potential Hallucinations - LLMs may occasionally print full data structures despite prompting
• Rate Limiting - External API constraints may require custom handling for large datasets