MCP Shield Pii

mcp盾牌

🛡️ MCP客户端和服务器的实时PII编辑代理——Python 3.12+的零延迟隐私，可选Python 3.14子解释器加速。

mcp盾牌 是位于MCP客户端（例如Claude Desktop）和任何下游MCP服务器之间的拦截网关代理。它在个人身份信息到达LLM的上下文窗口之前实时检测和屏蔽个人身份信息，确保GDPR/HIPAA符合单一 pip install.

为什么是mcp盾皮？

当AI代理从MCP服务器请求数据时，原始有效载荷（可能包含SSN、医疗记录或信用卡）直接流入LLM。组织可能面临超过数亿美元的GDPR/HIPAA罚款。 mcp盾牌 消除了协议层的这种风险。

┌──────────────┐     ┌─────────────────┐     ┌──────────────────┐
│ Claude       │────▶│ mcp-shield-pii  │────▶│ Downstream MCP   │
│ Desktop      │◀────│ (PII Redaction)  │◀────│ Server           │
└──────────────┘     └─────────────────┘     └──────────────────┘
                         ▲                           
                    PII masked before               
                    reaching the LLM                

安装

pip install mcp-shield-pii

对于基于NLP的检测（姓名、组织、地址）：

pip install mcp-shield-pii[nlp]
python -m spacy download en_core_web_sm

快速开始

1.扫描文本以查找PII

# Simple scan
mcp-shield-pii scan "Contact john@example.com, SSN 123-45-6789"

# JSON output
mcp-shield-pii scan --json "Patient MRN-123456 at 192.168.1.1"

# Different masking strategies
mcp-shield-pii scan --strategy partial "Card: 4111-1111-1111-1111"
mcp-shield-pii scan --strategy hash "Email: secret@corp.com"
mcp-shield-pii scan --strategy pseudo "Call 555-123-4567"

2.启动代理

# Basic proxy (stdio transport)
mcp-shield-pii proxy --downstream "npx -y @modelcontextprotocol/server-postgres postgresql://localhost/mydb"

# With config file
mcp-shield-pii proxy --downstream "python my_server.py" --config shield.toml

# Dry-run mode (log detections, don't modify payloads)
mcp-shield-pii proxy --downstream "npx my-mcp-server" --dry-run

3.克劳德桌面集成

添加到您的 claude_desktop_config.json:

{
  "mcpServers": {
    "my-server-shielded": {
      "command": "mcp-shield-pii",
      "args": [
        "proxy",
        "--downstream", "npx -y @modelcontextprotocol/server-postgres postgresql://localhost/mydb",
        "--config", "/path/to/shield.toml"
      ]
    }
  }
}

4.生成配置文件

mcp-shield-pii generate-config --output shield.toml

5.生成合规报告

mcp-shield-pii report --format markdown --output compliance_report.md

6.启动仪表板

mcp-shield-pii dashboard --port 8765
# Open http://127.0.0.1:8765

特性

v1.0--核心

v1.1--硬化

v2.0--企业版

检测到的实体类型

基于正则表达式（18种类型）

基于NLP（5种类型，需要 [nlp] 额外）

配置（shield.toml）

[shield]
default_masking_strategy = "redact"
default_confidence_threshold = 0.7
dry_run = false

[detection]
enable_regex = true
enable_nlp = false
enable_context_scoring = true

[entities.SSN]
masking_strategy = "redact"
confidence_threshold = 0.8

[entities.EMAIL]
masking_strategy = "pseudo"
confidence_threshold = 0.7

[tools.trusted_internal_tool]
action = "skip"

[tools.patient_records_api]
action = "strict"
masking_strategy = "redact"

[[webhooks]]
url = "https://hooks.slack.com/services/YOUR/WEBHOOK"
events = ["high_severity"]

[dashboard]
enabled = true
port = 8765

[metrics]
enabled = true
port = 9090

程序化API

from mcp_shield_pii.detection.regex_engine import RegexDetectionEngine
from mcp_shield_pii.masking.strategies import get_strategy
from mcp_shield_pii.pipeline import ShieldPipeline
from mcp_shield_pii.config.loader import ShieldConfig

# Simple detection
engine = RegexDetectionEngine()
results = engine.detect("Email john@corp.com, SSN 123-45-6789")
for r in results:
    print(f"{r.entity_type.value}: '{r.text}' (confidence: {r.confidence:.0%})")

# Full pipeline
config = ShieldConfig(default_masking_strategy="partial")
pipeline = ShieldPipeline(config)
masked, summary = pipeline.process_text("Contact admin@secret.org, card 4111-1111-1111-1111")
print(masked)  # "Contact a***@***.org, card ****-****-****-1111"
pipeline.close()

# Pseudo-anonymization
config = ShieldConfig(default_masking_strategy="pseudo")
pipeline = ShieldPipeline(config)
masked, _ = pipeline.process_text("Email alice@corp.com then alice@corp.com again")
print(masked)  # Same fake email both times (consistent mapping)
pipeline.close()

建筑

src/mcp_shield_pii/
├── __init__.py          # Public API exports
├── cli.py               # Typer CLI (6 commands)
├── pipeline.py          # Orchestration: detect → score → filter → mask → audit
├── compliance.py        # GDPR/HIPAA report generator
├── webhooks.py          # Async webhook alerts
├── detection/
│   ├── base.py          # EntityType enum, DetectionResult, protocols
│   ├── regex_engine.py  # 18 regex patterns + Luhn/IBAN validation
│   ├── nlp_engine.py    # spaCy NER detection (optional)
│   └── context_scorer.py # Context-aware confidence adjustment
├── masking/
│   ├── strategies.py    # Redact, partial, hash, pseudo-anonymization
│   └── reversible.py    # AES-256 Fernet reversible redaction
├── config/
│   ├── loader.py        # TOML config parser
│   └── watcher.py       # Hot-reload file watcher
├── proxy/
│   ├── __init__.py      # MCP JSON-RPC interceptor
│   └── stdio_proxy.py   # Bidirectional stdio transport
├── concurrency/
│   └── __init__.py      # Subinterpreter pool + ProcessPool fallback
├── metrics/
│   └── __init__.py      # Prometheus metrics + HTTP server
├── audit/
│   └── __init__.py      # JSONL audit logger
└── dashboard/
    └── __init__.py      # Web UI + REST API

贡献

看 贡献.md

许可证

麻省理工学院——见 许可证 了解详情。