广告技术MCP套件
一套10个用于广告技术的生产级模型上下文协议（MCP）服务器，每个服务器都具有分层多代理架构和高级ML算法。
建筑
每个MCP都遵循相同的模式：
Orchestrator Agent
├── Specialist Agents (domain experts)
│ ├── ML models (RF, GBM, Isolation Forest, etc.)
│ └── LLM reasoning layer
├── Critic Agent (adversarial review)
└── Ensemble Engine (weighted aggregation)
Shared Safety Layer
├── API Security (rate limiting, injection detection)
├── Design Validator (dark patterns, WCAG, budget guardrails)
└── Privacy Guard (PII detection, GDPR compliance)10 MCP
| # | MCP | 关键算法 | 关键代理 |
|---|---|---|---|
| 1 | 创新智慧 | 随机森林、孤立森林、指数衰减、CUSUM | 创意、观众、竞价、疲劳、简报、评论家、编曲 |
| 2 | 归因 | 马尔可夫链去除效应，岭MMM | 归因，批评 |
| 3 | 品牌安全 | TF-IDF+朴素贝叶斯 | 内容分类器，安全批评 |
| 4 | 有竞争力的英特尔 | KMeans+TF-IDF | 竞争对手英特尔，竞争评论家 |
| 5 | 受众数据 | 逻辑回归相似，Jaccard重叠 | 受众数据，受众评论家 |
| 6 | RTB优化器 | 梯度增强、汤普森采样 | RTB优化器、RTBCritic |
| 7 | 隐私合规 | 模式匹配、IAB TCF | 同意合规、合规批评 |
| 8 | 测量 | 频率学家+贝叶斯A/B，IQR/Z评分异常 | 测量，统计批评 |
| 9 | 网红 | 参与度分析，TF-IDF余弦相似性 | 影响力评分，影响力评分 |
| 10 | 创造性生产 | 蒙特卡洛功率分析，决策树 | 创意生产，生产评论家 |
设置
# 1. Install dependencies
pip install anthropic mcp scikit-learn numpy pandas scipy rich python-dotenv
# 2. Set API key
cp shared/config.py 0X-mcp-name/config.py # or set ANTHROPIC_API_KEY env var
# 3. Run any MCP server
python 01-creative-intelligence/mcp_server.py
python 02-attribution/mcp_server.py
# etc.
# 4. Or run the full demo (MCP 1)
python 01-creative-intelligence/run_demo.py添加到Claude桌面
将任何服务器添加到 claude_desktop_config.json:
{
"mcpServers": {
"adtech-creative": {
"command": "python",
"args": ["/path/to/01-creative-intelligence/mcp_server.py"],
"env": { "ANTHROPIC_API_KEY": "your-key" }
},
"adtech-attribution": {
"command": "python",
"args": ["/path/to/02-attribution/mcp_server.py"],
"env": { "ANTHROPIC_API_KEY": "your-key" }
}
}
}安全结构
这 shared/safety_agents/ 目录包含在每次工具调用时运行的代理：
api_security.py--速率限制、SQL/提示注入检测、PII编辑、审计日志记录design_validator.py--暗模式检测、WCAG可访问性、预算护栏
MCP 1深潜：创造性智力
旗舰MCP使用5级分层管道：
Stage 1: Specialist Agents (run in parallel)
├── Creative Performance Agent → Random Forest (200 trees, 5-fold CV)
├── Audience Intelligence Agent → Efficiency scoring (CTR × log(Reach))
├── Bidding Dynamics Agent → Win-rate/competition analysis
└── Fatigue Detection Agent → Exp decay + CUSUM change-point
Stage 2: Brief Agent → synthesises specialist outputs
Stage 3: Critic Agent → Isolation Forest + adversarial LLM review
Stage 4: Ensemble Engine → confidence-weighted soft voting
Stage 5: Orchestrator → final executive recommendation许可证
麻省理工学院