Step 6 MCP + AI エージェント連携

# グローバルにインストール（推奨）
npm install -g @brightdata/mcp-server
 
# 起動テスト
brightdata-mcp --version

pip install brightdata-mcp
 
# 起動テスト
brightdata-mcp --version

{
  "tools": [
    {
      "name": "brightdata-mcp",
      "type": "http",
      "description": "Bright Data MCP Server を介した Web 検索・スクレイピング",
      "endpoint": "http://localhost:8080/v1",
      "auth": {
        "type": "apiKey",
        "keyName": "X-API-KEY",
        "keyValue": "${BRIGHTDATA_API_KEY}"
      }
    }
  ]
}

{
  "name": "brightdata-mcp",
  "type": "http",
  "endpoint": "http://localhost:8080/v1",
  "auth": {
    "type": "apiKey",
    "header": "X-API-KEY",
    "value": "${BRIGHTDATA_API_KEY}"
  },
  "description": "MCP Server 経由で Web 検索・スクレイピング"
}

curl -X POST "http://localhost:8080/v1/search" \
  -H "Content-Type: application/json" \
  -H "X-API-KEY: $BRIGHTDATA_API_KEY" \
  -d '{
        "query": "Bright Data MCP 使い方",
        "engine": "google",
        "num_results": 3
      }'

{
  "results": [
    {
      "title": "Bright Data MCP Overview",
      "url": "https://docs.brightdata.com/ai/mcp-server/overview",
      "snippet": "MCP (Multi‑Channel Proxy) ...",
      "cite": "docs.brightdata.com"
    },
    ...
  ]
}

pip install langchain openai tiktoken

import os
import json
import requests
from langchain.llms import OpenAI
from langchain.prompts import PromptTemplate
from langchain.chains import LLMChain
 
# 環境変数
API_KEY = os.getenv("BRIGHTDATA_API_KEY")
MCP_ENDPOINT = "http://localhost:8080/v1/search"
 
def mcp_search(query: str, num_results: int = 3):
    payload = {
        "query": query,
        "engine": "google",
        "num_results": num_results
    }
    headers = {
        "Content-Type": "application/json",
        "X-API-KEY": API_KEY
    }
    resp = requests.post(MCP_ENDPOINT, headers=headers, json=payload, timeout=30)
    resp.raise_for_status()
    return resp.json()["results"]
 
# OpenAI LLM（Claude の代わりに OpenAI を使用例例）
llm = OpenAI(temperature=0, model_name="gpt-4o-mini")
 
# 要約プロンプト
summary_prompt = PromptTemplate(
    input_variables=["content"],
    template=(
        "以下のテキストを 150 文字以内で要約してください。\n"
        "テキスト:\n{content}"
    )
)
summary_chain = LLMChain(llm=llm, prompt=summary_prompt)
 
def answer_with_citation(question: str):
    # 1. 検索
    results = mcp_search(question)
    # 2. 取得したページの本文を簡易取得（ここでは snippet のみ使用）
    contents = "\n\n".join([f"{r['title']}:\n{r['snippet']}" for r in results])
    # 3. 要約
    summary = summary_chain.run(content=contents)
    # 4. 引用付き回答組み立て
    citations = "\n".join([f"[{i+1}] {r['url']}" for i, r in enumerate(results)])
    answer = f"{summary}\n\n引用:\n{citations}"
    return answer
 
if __name__ == "__main__":
    q = "Bright Data MCP を使ったスクレイピングのベストプラクティスは？"
    print(answer_with_citation(q))

export BRIGHTDATA_API_KEY=YOUR_API_KEY
python mcp_agent.py