GitHub - neul-labs/ormai: Give your AI agents database ac...

Give your AI agents database access without the risk.

OrmAI wraps your existing ORM models in a policy-enforced runtime. Your agents get typed tools for querying and writing data — while you keep control over what they can see and do. No raw SQL. No prompt injection into your database. Just safe, auditable, tenant-scoped database tools.

Available for Python and TypeScript/Node.js.

Why OrmAI?

Building AI agents that interact with your database? You have probably thought about:

"What if the agent reads sensitive data?" → Field-level policies hide or mask PII automatically.
"What if it runs wild queries?" → Query budgets and row limits prevent runaway costs.
"How do I audit what it did?" → Every operation is logged with full context.
"What about multi-tenant isolation?" → Tenant scoping is built-in, not bolted on.
"Which ORM do we use?" → Works with SQLAlchemy, Prisma, Drizzle, TypeORM, Tortoise, Django, SQLModel, and Peewee.

OrmAI solves these at the ORM layer — not the prompt layer.

Pick Your Stack

	Python	TypeScript / Node.js
Package	`pip install ormai`	`npm install @ormai/core`
ORMs	SQLAlchemy, Tortoise, Django, SQLModel, Peewee	Prisma, Drizzle, TypeORM
Integrations	OpenAI, LangChain, LlamaIndex, MCP, FastAPI	Vercel AI SDK, LangChain.js, OpenAI, Anthropic, LlamaIndex.ts, Mastra, MCP
Quickstart	`ormai.quickstart`	`@ormai/utils`
Docs	Python Guide	TS Guide

Python Quick Start

# With your ORM of choice
pip install ormai[sqlalchemy]
# or
pip install ormai[prisma]

from ormai.quickstart import mount_sqlalchemy
from ormai.utils import DEFAULT_DEV

# Your existing SQLAlchemy models + session
toolset = mount_sqlalchemy(
    engine=engine,
    session_factory=Session,
    policy=DEFAULT_DEV
)

# Done. Your agent now has: db.query, db.get, db.aggregate, db.describe_schema

TypeScript Quick Start

# Core (required)
npm install @ormai/core

# Choose your ORM adapter
npm install @ormai/prisma

import { PrismaClient } from '@prisma/client';
import { PrismaAdapter } from '@ormai/prisma';
import { PolicyBuilder, createContext } from '@ormai/core';
import { createGenericTools } from '@ormai/tools';

const prisma = new PrismaClient();
const adapter = new PrismaAdapter({ prisma });
const schema = await adapter.introspect();

const policy = new PolicyBuilder('prod')
  .registerModels(['Customer', 'Order'])
  .tenantScope('tenantId')
  .denyFields('*password*')
  .maskFields('*email*')
  .build();

const tools = createGenericTools({ adapter, policy, schema });

const ctx = createContext({
  tenantId: 'tenant-123',
  userId: 'user-456',
  db: prisma,
  roles: ['admin'],
});

// Your agent now has safe database tools
const result = await tools[0].execute({
  model: 'Order',
  where: [{ field: 'status', op: 'eq', value: 'pending' }],
  take: 10,
}, ctx);

What You Get Out of the Box

Feature	What It Does
Read-safe tools	`db.query`, `db.get`, `db.aggregate`, `db.describe_schema` — no raw SQL
Write-safe tools	`db.create`, `db.update`, `db.delete`, `db.bulk_update` — gated by policy
Field-level policies	Hide passwords, mask emails, deny sensitive columns automatically
Tenant scoping	`.tenantScope('tenant_id')` auto-filters every query per user
Query budgets	Max rows, max includes depth, statement timeouts per model
Audit logging	Every call logged with principal, tenant, trace ID, input, output
Human approval gates	Require reason or approval for writes on sensitive models
Schema introspection	Auto-discovers models, fields, relations, primary keys
Multi-framework	LangChain, OpenAI, Vercel AI SDK, LlamaIndex, Mastra, FastAPI, MCP

Architecture

┌─────────────────────────────────────────────────────────────┐
│                        Your Agent                           │
└──────────────────────────┬──────────────────────────────────┘
                           │ calls tools
┌──────────────────────────▼──────────────────────────────────┐
│                    OrmAI Runtime                            │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────┐│
│  │   Policy    │  │   Audit     │  │    Tenant Scope     ││
│  │  Enforcer   │  │   Logger    │  │      Filter         ││
│  └─────────────┘  └─────────────┘  └─────────────────────┘│
└──────────────────────────┬──────────────────────────────────┘
                           │ parameterized queries only
┌──────────────────────────▼──────────────────────────────────┐
│          Your ORM (SQLAlchemy / Prisma / Drizzle / ...)   │
└─────────────────────────────────────────────────────────────┘

OrmAI sits between your agent and your ORM. It compiles agent requests into type-safe ORM queries, enforces policies, logs everything, and returns structured results. Your database never sees raw SQL from the agent.

Documentation

docs.neullabs.com/ormai — Full guides, API reference, and examples.

Installation

Python

pip install ormai[sqlalchemy]
pip install ormai[tortoise]
pip install ormai[peewee]
pip install ormai[django]
pip install ormai[sqlmodel]

# Or all adapters
pip install ormai[all]

TypeScript / Node.js

# Core (required)
npm install @ormai/core

# ORM adapters
npm install @ormai/prisma
npm install @ormai/drizzle
npm install @ormai/typeorm

# Optional packages
npm install @ormai/tools     # Generic database tools
npm install @ormai/store     # Audit logging
npm install @ormai/mcp       # MCP server
npm install @ormai/integrations  # Framework adapters
npm install @ormai/utils     # PolicyBuilder and helpers

Policy Configuration

Python

from ormai.utils import PolicyBuilder, DEFAULT_PROD

policy = (
    PolicyBuilder(DEFAULT_PROD)
    .register_models([Customer, Order])
    .deny_fields("*password*", "*secret*", "*token*")
    .mask_fields(["email", "phone"])
    .tenant_scope("tenant_id")
    .enable_writes(models=["Order"], require_reason=True)
    .build()
)

TypeScript

import { PolicyBuilder } from '@ormai/core';

const policy = new PolicyBuilder('prod')
  .registerModels(['Customer', 'Order', 'Product'])
  .tenantScope('tenantId')
  .denyFields('*password*')
  .maskFields('*email*')
  .allowRelations('Order', ['customer', 'items'])
  .enableWrites(['Order'], {
    allowCreate: true,
    allowUpdate: true,
    allowDelete: false,
    maxAffectedRows: 10,
  })
  .defaultBudgetConfig({
    maxRows: 100,
    maxIncludesDepth: 2,
    statementTimeoutMs: 5000,
  })
  .build();

Presets: DEFAULT_DEV (permissive), DEFAULT_INTERNAL (moderate), DEFAULT_PROD (strict)

Agent Framework Integrations

Framework	Python Package	TypeScript Package
OpenAI	`ormai`	`@ormai/integrations`
LangChain	`ormai`	`@ormai/integrations`
Vercel AI SDK	—	`@ormai/integrations`
LlamaIndex	`ormai`	`@ormai/integrations`
Mastra	—	`@ormai/integrations`
Anthropic	—	`@ormai/integrations`
FastAPI	`ormai`	—
MCP	`ormai`	`@ormai/mcp`

Supported ORMs

ORM	Python	TypeScript
SQLAlchemy	✅	—
Prisma	—	✅
Drizzle	—	✅
TypeORM	—	✅
SQLModel	✅	—
Django ORM	✅	—
Tortoise ORM	✅	—
Peewee	✅	—

Benchmark: OrmAI vs Text-to-SQL

We benchmarked against the Spider dataset — 1034 natural language queries:

Metric	OrmAI	Text-to-SQL
SQL Injection possible	No	Yes
Unsafe ops executed	0	23
Full audit trail	Yes	No

# Try it yourself
pip install ormai[benchmark]
python examples/spider_demo.py run --limit 20

Examples

examples/spider_demo.py — Benchmark demo
examples/fastapi-sqlalchemy/ — FastAPI + SQLAlchemy integration
ormai-ts/ — TypeScript monorepo with Prisma, Drizzle, and TypeORM examples

More examples at docs.neullabs.com/ormai/examples.

Contributing

git clone https://github.com/neul-labs/ormai.git
cd ormai

# Python
uv sync --dev
uv run pytest

# TypeScript
cd ormai-ts
npm install
npm run build
npm run test

See contributing guide for development setup and guidelines.

推荐订阅源

Show HN

Why OrmAI?

Pick Your Stack

Python Quick Start

TypeScript Quick Start

What You Get Out of the Box

Architecture

Documentation

Installation

Python

TypeScript / Node.js

Policy Configuration

Python

TypeScript

Agent Framework Integrations

Supported ORMs

Benchmark: OrmAI vs Text-to-SQL

Examples

Contributing