EnrichMCP：为 Agent 优化的 Python ORM

面向 AI Agent 的 ORM——将你的数据模型转化为语义化 MCP 层。

EnrichMCP 是一个 Python 框架，帮助 AI Agent 理解并浏览你的数据。它基于 MCP（Model Context Protocol，模型上下文协议）构建，增加了一个语义层，将数据模型转化为具备类型、可发现的工具——就像专为 AI 打造的 ORM。

你可以把它理解为 AI Agent 领域的 SQLAlchemy。EnrichMCP 可以自动：

• 根据数据模型生成具备类型的工具
• 处理实体之间的关系（用户 → 订单 → 商品）
• 提供 schema 发现能力，让 AI Agent 理解你的数据结构
• 使用 Pydantic 模型验证所有输入和输出
• 兼容任意后端——数据库、API 或自定义逻辑

pip install enrichmcp

# With SQLAlchemy support
pip install enrichmcp[sqlalchemy]

方案一：我有 SQLAlchemy 模型（30 秒）

将现有的 SQLAlchemy 模型转化为 AI 可浏览的 API：

from enrichmcp import EnrichMCP
from enrichmcp.sqlalchemy import (
    include_sqlalchemy_models,
    sqlalchemy_lifespan,
    EnrichSQLAlchemyMixin,
)
from sqlalchemy import ForeignKey
from sqlalchemy.ext.asyncio import create_async_engine
from sqlalchemy.orm import DeclarativeBase, Mapped, mapped_column, relationship

engine = create_async_engine("postgresql+asyncpg://user:pass@localhost/db")

# Add the mixin to your declarative base
class Base(DeclarativeBase, EnrichSQLAlchemyMixin):
    pass

class User(Base):
    """User account."""

    __tablename__ = "users"

    id: Mapped[int] = mapped_column(primary_key=True, info={"description": "Unique user ID"})
    email: Mapped[str] = mapped_column(unique=True, info={"description": "Email address"})
    status: Mapped[str] = mapped_column(default="active", info={"description": "Account status"})
    orders: Mapped[list["Order"]] = relationship(
        back_populates="user", info={"description": "All orders for this user"}
    )

class Order(Base):
    """Customer order."""

    __tablename__ = "orders"

    id: Mapped[int] = mapped_column(primary_key=True, info={"description": "Order ID"})
    user_id: Mapped[int] = mapped_column(
        ForeignKey("users.id"), info={"description": "Owner user ID"}
    )
    total: Mapped[float] = mapped_column(info={"description": "Order total"})
    user: Mapped[User] = relationship(
        back_populates="orders", info={"description": "User who placed the order"}
    )

# That's it! Create your MCP app
app = EnrichMCP(
    "E-commerce Data",
    "API generated from SQLAlchemy models",
    lifespan=sqlalchemy_lifespan(Base, engine, cleanup_db_file=True),
)
include_sqlalchemy_models(app, Base)

if __name__ == "__main__":
    app.run()

• explore_data_model()——了解完整的 schema
• list_users(status='active')——使用过滤条件查询
• get_user(id=123)——获取指定记录
• 浏览关系：user.orders → order.user

方案二：我有 REST API（2 分钟）

为现有 API 包装一层语义理解能力：

from typing import Literal
from enrichmcp import EnrichMCP, EnrichModel, Relationship
from pydantic import Field
import httpx

app = EnrichMCP("API Gateway", "Wrapper around existing REST APIs")
http = httpx.AsyncClient(base_url="https://api.example.com")

@app.entity()
class Customer(EnrichModel):
    """Customer in our CRM system."""

    id: int = Field(description="Unique customer ID")
    email: str = Field(description="Primary contact email")
    tier: Literal["free", "pro", "enterprise"] = Field(description="Subscription tier")

    # Define navigable relationships
    orders: list["Order"] = Relationship(description="Customer's purchase history")

@app.entity()
class Order(EnrichModel):
    """Customer order from our e-commerce platform."""

    id: int = Field(description="Order ID")
    customer_id: int = Field(description="Associated customer")
    total: float = Field(description="Order total in USD")
    status: Literal["pending", "shipped", "delivered"] = Field(description="Order status")

    customer: Customer = Relationship(description="Customer who placed this order")

# Define how to fetch data
@app.retrieve()
async def get_customer(customer_id: int) -> Customer:
    """Fetch customer from CRM API."""
    response = await http.get(f"/api/customers/{customer_id}")
    return Customer(**response.json())

# Define relationship resolvers
@Customer.orders.resolver
async def get_customer_orders(customer_id: int) -> list[Order]:
    """Fetch orders for a customer."""
    response = await http.get(f"/api/customers/{customer_id}/orders")
    return [Order(**order) for order in response.json()]

@Order.customer.resolver
async def get_order_customer(order_id: int) -> Customer:
    """Fetch the customer for an order."""
    response = await http.get(f"/api/orders/{order_id}/customer")
    return Customer(**response.json())

app.run()

方案三：我想要完全掌控（5 分钟）

使用自定义逻辑构建完整的数据层：

from enrichmcp import EnrichMCP, EnrichModel, Relationship
from datetime import datetime
from decimal import Decimal
from pydantic import Field

app = EnrichMCP("Analytics Platform", "Custom analytics API")

db = ...  # your database connection

@app.entity()
class User(EnrichModel):
    """User with computed analytics fields."""

    id: int = Field(description="User ID")
    email: str = Field(description="Contact email")
    created_at: datetime = Field(description="Registration date")

    # Computed fields
    lifetime_value: Decimal = Field(description="Total revenue from user")
    churn_risk: float = Field(description="ML-predicted churn probability 0-1")

    # Relationships
    orders: list["Order"] = Relationship(description="Purchase history")
    segments: list["Segment"] = Relationship(description="Marketing segments")

@app.entity()
class Segment(EnrichModel):
    """Dynamic user segment for marketing."""

    name: str = Field(description="Segment name")
    criteria: dict = Field(description="Segment criteria")
    users: list[User] = Relationship(description="Users in this segment")

@app.entity()
class Order(EnrichModel):
    """Simplified order record."""

    id: int = Field(description="Order ID")
    user_id: int = Field(description="Owner user ID")
    total: Decimal = Field(description="Order total")

@User.orders.resolver
async def list_user_orders(user_id: int) -> list[Order]:
    """Fetch orders for a user."""
    rows = await db.query(
        "SELECT * FROM orders WHERE user_id = ? ORDER BY id DESC",
        user_id,
    )
    return [Order(**row) for row in rows]

@User.segments.resolver
async def list_user_segments(user_id: int) -> list[Segment]:
    """Fetch segments that include the user."""
    rows = await db.query(
        "SELECT s.* FROM segments s JOIN user_segments us ON s.name = us.segment_name WHERE us.user_id = ?",
        user_id,
    )
    return [Segment(**row) for row in rows]

@Segment.users.resolver
async def list_segment_users(name: str) -> list[User]:
    """List users in a segment."""
    rows = await db.query(
        "SELECT u.* FROM users u JOIN user_segments us ON u.id = us.user_id WHERE us.segment_name = ?",
        name,
    )
    return [User(**row) for row in rows]

# Complex resource with business logic
@app.retrieve()
async def find_high_value_at_risk_users(
    lifetime_value_min: Decimal = 1000, churn_risk_min: float = 0.7, limit: int = 100
) -> list[User]:
    """Find valuable customers likely to churn."""
    users = await db.query(
        """
        SELECT * FROM users
        WHERE lifetime_value >= ? AND churn_risk >= ?
        ORDER BY lifetime_value DESC
        LIMIT ?
        """,
        lifetime_value_min,
        churn_risk_min,
        limit,
    )
    return [User(**u) for u in users]

# Async computed field resolver
@User.lifetime_value.resolver
async def calculate_lifetime_value(user_id: int) -> Decimal:
    """Calculate total revenue from user's orders."""
    total = await db.query_single("SELECT SUM(total) FROM orders WHERE user_id = ?", user_id)
    return Decimal(str(total or 0))

# ML-powered field
@User.churn_risk.resolver
async def predict_churn_risk(user_id: int) -> float:
    """Run churn prediction model."""
    ctx = app.get_context()
    features = await gather_user_features(user_id)
    model = ctx.get("ml_models")["churn"]
    return float(model.predict_proba(features)[0][1])

app.run()

🔍 自动发现 Schema

AI Agent 只需一次调用，即可探索完整的数据模型：

schema = await explore_data_model()
# Returns complete schema with entities, fields, types, and relationships

🔗 关系浏览

关系只需定义一次，AI Agent 就能自然地沿关系进行浏览：

# AI can navigate: user → orders → products → categories
user = await get_user(123)
orders = await user.orders()  # Automatic resolver
products = await orders[0].products()

🛡️ 类型安全与验证

每次交互都会经过完整的 Pydantic 验证：

@app.entity()
class Order(EnrichModel):
    total: float = Field(ge=0, description="Must be positive")
    email: EmailStr = Field(description="Customer email")
    status: Literal["pending", "shipped", "delivered"]

describe_model() 会列出这些允许值，让 Agent 知道哪些选项有效。

字段默认不可变。你可以将字段标记为可变，并使用自动生成的 patch 模型进行更新：

@app.entity()
class Customer(EnrichModel):
    id: int = Field(description="ID")
    email: str = Field(json_schema_extra={"mutable": True}, description="Email")

@app.create()
async def create_customer(email: str) -> Customer: ...

@app.update()
async def update_customer(cid: int, patch: Customer.PatchModel) -> Customer: ...

@app.delete()
async def delete_customer(cid: int) -> bool: ...

📄 内置分页

优雅处理大型数据集：

from enrichmcp import PageResult

@app.retrieve()
async def list_orders(page: int = 1, page_size: int = 50) -> PageResult[Order]:
    orders, total = await db.get_orders_page(page, page_size)
    return PageResult.create(items=orders, page=page, page_size=page_size, total_items=total)

更多示例请参阅 Pagination Guide。

🔐 上下文与身份验证

传递身份验证信息、数据库连接或任意上下文：

from pydantic import Field
from enrichmcp import EnrichModel

class UserProfile(EnrichModel):
    """User profile information."""

    user_id: int = Field(description="User ID")
    bio: str | None = Field(default=None, description="Short bio")

@app.retrieve()
async def get_user_profile(user_id: int) -> UserProfile:
    ctx = app.get_context()
    # Access context provided by MCP client
    auth_user = ctx.get("authenticated_user_id")
    if auth_user != user_id:
        raise PermissionError("Can only access your own profile")
    return await db.get_profile(user_id)

将结果保存在按请求、按用户或全局划分的缓存中，以减少 API 开销：

@app.retrieve()
async def get_customer(cid: int) -> Customer:
    ctx = app.get_context()

    async def fetch() -> Customer:
        return await db.get_customer(cid)

    return await ctx.cache.get_or_set(f"customer:{cid}", fetch)

使用 EnrichParameter 为工具参数提供示例和元数据：

from enrichmcp import EnrichParameter

@app.retrieve()
async def greet_user(name: str = EnrichParameter(description="user name", examples=["bob"])) -> str:
    return f"Hello {name}"

工具描述中会包含参数类型、描述和示例。

🌐 HTTP 与 SSE 支持

通过标准输出（默认）、SSE 或 HTTP 提供 API：

app.run()  # stdio default
app.run(transport="streamable-http")

EnrichMCP 在 MCP 之上增加了三个关键层：

• 语义层——AI Agent 不仅能理解数据结构，还能理解数据的含义
• 数据层——具备验证和关系能力的类型安全模型
• 控制层——身份验证、分页和业务逻辑

最终效果是：AI Agent 可以像开发者使用 ORM 一样自然地处理你的数据。

服务端 LLM Sampling

EnrichMCP 可以通过 MCP 的 sampling 功能请求语言模型补全。在任意资源中调用 ctx.ask_llm() 或其别名 ctx.sampling()，已连接的客户端就会选择一个 LLM 并支付相关用量费用。你可以使用 model_preferences、allow_tools 和 max_tokens 等选项调整其行为。更多详情请参阅 docs/server_side_llm.md。

查看 examples 目录中的示例：

• hello_world——最精简的 EnrichMCP 应用
• hello_world_http——使用 streamable HTTP 的 HTTP 示例
• shop_api——支持分页和过滤的内存版商店 API
• shop_api_sqlite——由 SQLite 提供支持的版本
• shop_api_gateway——将 EnrichMCP 用作 FastAPI 前端的网关
• sqlalchemy_shop——根据 SQLAlchemy 模型自动生成 API
• mutable_crud——演示可变字段和 CRUD decorators
• caching——演示 ContextCache 的用法
• basic_memory——使用 FileMemoryStore 构建的简单笔记 API
• openai_chat_agent——用于 MCP 示例的交互式聊天客户端

📖 完整文档

🚀 入门指南

欢迎贡献！详情请参阅 CONTRIBUTING.md。

该仓库要求使用 Python 3.11 或更高版本。Makefile 中包含用于创建虚拟环境和运行测试的命令：

make setup            # create .venv and install dependencies
source .venv/bin/activate
make test             # run the test suite

这会安装所有开发 extras 和 pre-commit hooks，因此 make lint 或 make docs 等命令可以立即使用。

Apache 2.0——请参阅 LICENSE。

由 Featureform 构建 • MCP Protocol