内容
# EnrichMCP
**The ORM for AI Agents - Turn your data model into a semantic MCP layer**
[](https://github.com/featureform/enrichmcp/actions/workflows/ci.yml)
[](https://codecov.io/gh/featureform/enrichmcp)
[](https://pypi.org/project/enrichmcp/)
[](https://www.python.org/downloads/)
[](https://github.com/featureform/enrichmcp/blob/main/LICENSE)
[](https://featureform.github.io/enrichmcp)
EnrichMCP is a Python framework that helps AI agents understand and navigate your data. Built on MCP (Model Context Protocol), it adds a semantic layer that turns your data model into typed, discoverable tools - like an ORM for AI.
## What is EnrichMCP?
Think of it as SQLAlchemy for AI agents. EnrichMCP automatically:
- **Generates typed tools** from your data models
- **Handles relationships** between entities (users → orders → products)
- **Provides schema discovery** so AI agents understand your data structure
- **Validates all inputs/outputs** with Pydantic models
- **Works with any backend** - databases, APIs, or custom logic
## Installation
```bash
pip install enrichmcp
# With SQLAlchemy support
pip install enrichmcp[sqlalchemy]
```
## Show Me Code
### Option 1: I Have SQLAlchemy Models (30 seconds)
Transform your existing SQLAlchemy models into an AI-navigable API:
```python
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()
```
AI agents can now:
- `explore_data_model()` - understand your entire schema
- `list_users(status='active')` - query with filters
- `get_user(id=123)` - fetch specific records
- Navigate relationships: `user.orders` → `order.user`
### Option 2: I Have REST APIs (2 minutes)
Wrap your existing APIs with semantic understanding:
```python
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()
```
### Option 3: I Want Full Control (5 minutes)
Build a complete data layer with custom logic:
```python
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()
```
## Key Features
### 🔍 Automatic Schema Discovery
AI agents explore your entire data model with one call:
```python
schema = await explore_data_model()
# Returns complete schema with entities, fields, types, and relationships
```
### 🔗 Relationship Navigation
Define relationships once, AI agents traverse naturally:
```python
# AI can navigate: user → orders → products → categories
user = await get_user(123)
orders = await user.orders() # Automatic resolver
products = await orders[0].products()
```
### 🛡️ Type Safety & Validation
Full Pydantic validation on every interaction:
```python
@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()` will list these allowed values so agents know the valid options.
### ✏️ Mutability & CRUD
Fields are immutable by default. Mark them as mutable and use
auto-generated patch models for updates:
```python
@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: ...
```
### 📄 Pagination Built-in
Handle large datasets elegantly:
```python
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)
```
See the [Pagination Guide](https://featureform.github.io/enrichmcp/pagination) for more examples.
### 🔐 Context & Authentication
Pass auth, database connections, or any context:
```python
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)
```
### ⚡ Request Caching
Reduce API overhead by storing results in a per-request, per-user, or global cache:
```python
@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)
```
### 🧭 Parameter Hints
Provide examples and metadata for tool parameters using `EnrichParameter`:
```python
from enrichmcp import EnrichParameter
@app.retrieve()
async def greet_user(name: str = EnrichParameter(description="user name", examples=["bob"])) -> str:
return f"Hello {name}"
```
Tool descriptions will include the parameter type, description, and examples.
### 🌐 HTTP & SSE Support
Serve your API over standard output (default), SSE, or HTTP:
```python
app.run() # stdio default
app.run(transport="streamable-http")
```
## Why EnrichMCP?
EnrichMCP adds three critical layers on top of MCP:
1. **Semantic Layer** - AI agents understand what your data means, not just its structure
2. **Data Layer** - Type-safe models with validation and relationships
3. **Control Layer** - Authentication, pagination, and business logic
The result: AI agents can work with your data as naturally as a developer using an ORM.
## Server-Side LLM Sampling
EnrichMCP can request language model completions through MCP's **sampling**
feature. Call `ctx.ask_llm()` or the `ctx.sampling()` alias from any resource
and the connected client will choose an LLM and pay for the usage. You can tune
behavior using options like `model_preferences`, `allow_tools`, and
`max_tokens`. See [docs/server_side_llm.md](docs/server_side_llm.md) for more
details.
## Examples
Check out the [examples directory](examples/README.md):
- [hello_world](examples/hello_world) - The smallest possible EnrichMCP app
- [hello_world_http](examples/hello_world_http) - HTTP example using streamable HTTP
- [shop_api](examples/shop_api) - In-memory shop API with pagination and filters
- [shop_api_sqlite](examples/shop_api_sqlite) - SQLite-backed version
- [shop_api_gateway](examples/shop_api_gateway) - EnrichMCP as a gateway in front of FastAPI
- [sqlalchemy_shop](examples/sqlalchemy_shop) - Auto-generated API from SQLAlchemy models
- [mutable_crud](examples/mutable_crud) - Demonstrates mutable fields and CRUD decorators
- [caching](examples/caching) - Demonstrates ContextCache usage
- [basic_memory](examples/basic_memory) - Simple note-taking API using FileMemoryStore
- [openai_chat_agent](examples/openai_chat_agent) - Interactive chat client for MCP examples
## Documentation
- 📖 [Full Documentation](https://featureform.github.io/enrichmcp)
- 🚀 [Getting Started Guide](https://featureform.github.io/enrichmcp/getting-started)
- 🔧 [API Reference](https://featureform.github.io/enrichmcp/api)
## Contributing
We welcome contributions! See [CONTRIBUTING.md](CONTRIBUTING.md) for details.
## Development Setup
The repository requires **Python 3.11** or newer. The Makefile includes
commands to create a virtual environment and run the tests:
```bash
make setup # create .venv and install dependencies
source .venv/bin/activate
make test # run the test suite
```
This installs all development extras and pre-commit hooks so commands like
`make lint` or `make docs` work right away.
## License
Apache 2.0 - See [LICENSE](LICENSE)
---
Built by [Featureform](https://featureform.com) • [MCP Protocol](https://modelcontextprotocol.io)
