此寄存器包含 **O’Reilly Demo Day ** 会话的代码和多克设置 :

奥赖利·德莫日* ** 用MCP建立强有力的AI应用程序和代理人**

– – . . .

□ 德莫概览

这个代码显示了自主代理如何可以使用MCP不仅连接工具,而且还可以相互沟通.

我们运行一个多代理编码系统,其中代理商通过:

• 生成Python解决方案,
• 在沙盒内测试和确定基准,
• 根据结构化反馈加以改进,
• 并一起发展代码 提高质量。

MCP作为推理代理人之间的共同协议,能够结构化的协调、思考和持续创新。 这是走向自我改进人工智能系统的一个步骤。

□为什么是MCP?

传统的地方行刑(如Python REPL)带有风险和局限性:

• ** 安全风险**: 执行主机上任意生成的代码可能导致数据泄露或系统损坏.
• ** 孤零零零**:错误(如无限回路,回路错误)可以挂起过程.
• ** 难以缩放**: 跨节点运行多个代理的代码需要管弦和沙箱.

MCP + MCTS Coding Agents

This repository contains the code and Docker setup for the O’Reilly Demo Day session:

O’Reilly Demo Day
Building Robust AI Apps and Agents with MCP

---

Demo Overview

This code shows how autonomous agents can use MCP not just to connect with tools but to communicate with each other.

We run a multi-agent coding system where agents collaborate by:

• generating Python solutions,
• testing and benchmarking them inside a sandbox,
• refining based on structured feedback,
• and evolving code together toward higher quality.

MCP acts as the shared protocol between reasoning agents, enabling structured coordination, reflection, and continuous innovation.
This is a step toward self-improving AI systems.

Why MCP?

Traditional local execution (like a Python REPL) comes with risks and limitations:

• Security risks: Executing arbitrary generated code on the host machine can lead to data leaks or system damage.
• Lack of isolation: Mistakes (e.g. infinite loops, recursion errors) can hang the process.
• Difficult scaling: Running code on multiple agents across nodes requires orchestration and sandboxing.

MCP (Model Context Protocol) solves these problems:

• Runs Python in Pyodide + Deno, sandboxed from the host OS.
• Automatically installs packages into the sandbox when needed.
• Captures stdout, stderr, and return values cleanly.
• Works across nodes and clusters (e.g. Kubernetes with gVisor).
• Provides a shared protocol for inter-agent communication, not just tool calls.

This makes MCP an excellent fit for multi-agent systems where reasoning agents collaborate safely.

---

Prerequisites

• Docker installed (version 24+ recommended).
• An OpenAI API key stored in a .env file.

Your .env file should look like this:

OPENAI_API_KEY=sk-...
OPENAI_MODEL=gpt-4o

---

Step 1: Build Docker image

The provided Dockerfile builds a container with:

• Python 3.11 slim
• Deno (for MCP Pyodide backend)
• mcp-run-python, treequest, langgraph, langchain-openai, and other dependencies

Build the image:

docker build -t mcp-run-python:latest .

Step 2: Start the MCP server (stdio mode)

This runs the Python MCP server inside the container:

docker run --rm -it mcp-run-python:latest /usr/local/bin/entrypoint.sh stdio

Step 3: Run the Tree Search + MCP Demo

From the project root, run:

docker run --rm -it \
  --env-file .env \
  -v "$PWD":/app -w /app \
  mcp-run-python:latest \
  python treesearch_fib.py

This executes the LangGraph + TreeQuest + MCP sandboxed agents pipeline. The system will:

• Iteratively generate Fibonacci implementations,
• Run unit tests and benchmarks in the sandbox,
• Refine the code based on structured feedback,
• Track performance improvements step by step,