⚠️ Early Alpha — Under Heavy Development

NinjaStack is actively undergoing heavy development and is not yet ready for production use. APIs, schemas, and architecture may change without notice. We're still in early alpha and welcome new contributors and anyone willing to explore a new frontier of agentic backend architecture. If that sounds like your kind of thing — jump in, open issues, and help shape what this becomes.

🥷 NinjaStack

Schema-first agentic backend framework.
Point at a database, get a full agentic backend with AI agents, GraphQL, auth, and UI.

Homepage · Documentation · Examples

---

What is NinjaStack?

NinjaStack transforms database schemas into fully functional agentic backends. Define your data model once — through database introspection or conversational design — and the framework generates AI agents, GraphQL APIs, authentication, RBAC, and deployment manifests.

# Connect to your database, discover the schema
ninjastack introspect --db postgres://localhost/myapp

# Generate everything: models, agents, GraphQL, auth
ninjastack sync

# Run your agentic backend
ninjastack serve
# → Agentic backend at http://localhost:8000
# → GraphQL playground at /graphql
# → Agent chat at /chat

No database yet? Chat with the AI setup assistant to design your schema through natural dialogue:

ninjastack init --interactive
# "I need a bookstore with books, customers, orders, and reviews..."

Key Features

Feature	Description
🔍 Database Introspection	Auto-discover entities from PostgreSQL, MongoDB, Neo4j, or vector stores
🤖 ADK Agent Generation	Google ADK agents with scoped CRUD tools per entity
🧬 Agentic Schema Definition	Typed, composable schema language — your single source of truth
🔐 Auth & RBAC	Pluggable auth (OAuth2, JWT, API keys) with declarative role-based permissions
📊 GraphQL Generation	Strawberry types, queries, and mutations from schema
💬 Conversational Setup	Design your schema through natural dialogue with Gemini
🎯 Tool Scoping	Each agent only sees its own tools — no leaking across boundaries
🚀 K8s Deployment	Helm charts and manifests generated automatically
🔄 Polyglot Persistence	Unified layer across SQL, NoSQL, graph, and vector databases

Architecture

NinjaStack organizes agents in a three-tier hierarchy with explicit ownership at every level:

graph TD
    C["🎯 Coordinator Agent<br/><small>LLM · gemini-2.5-pro · Intent routing</small>"]
    C --> D1["📚 Catalog Domain<br/><small>gemini-2.5-flash · Medium reasoning</small>"]
    C --> D2["🛒 Commerce Domain<br/><small>gemini-2.5-pro · High reasoning</small>"]
    D1 --> B["📖 Book Agent"]
    D1 --> R["⭐ Review Agent"]
    D2 --> Cu["👤 Customer Agent"]
    D2 --> O["📦 Order Agent"]
    B --> P["🗄️ Unified Persistence Layer<br/><small>SQL · MongoDB · Neo4j · ChromaDB</small>"]
    R --> P
    Cu --> P
    O --> P

    style C fill:#166534,color:#fff,stroke:#22c55e
    style D1 fill:#1e3a5f,color:#fff,stroke:#3b82f6
    style D2 fill:#1e3a5f,color:#fff,stroke:#3b82f6
    style B fill:#854d0e,color:#fff,stroke:#eab308
    style R fill:#854d0e,color:#fff,stroke:#eab308
    style Cu fill:#854d0e,color:#fff,stroke:#eab308
    style O fill:#854d0e,color:#fff,stroke:#eab308
    style P fill:#581c87,color:#fff,stroke:#a855f7

• Data Agents — Deterministic CRUD. No LLM. One entity, scoped tools. Fast and testable.
• Domain Agents — LLM-powered. Own a business domain. Delegate to data agents. Configurable reasoning.
• Coordinator — Top-level router. Classifies intent. Synthesizes cross-domain results.

📚 Full architecture docs →

Quick Start

Prerequisites

• Python 3.12+
• uv package manager

Install from source

git clone https://github.com/codeninja/ninja-stack.git
cd ninja-stack
uv sync

Run the examples

All examples use a bookstore domain and work without any API keys:

# Schema definition
PYTHONPATH=examples/bookstore uv run python examples/bookstore/01_schema_definition.py

# Data agents (deterministic CRUD)
PYTHONPATH=examples/bookstore uv run python examples/bookstore/03_data_agents.py

# Full end-to-end pipeline
PYTHONPATH=examples/bookstore uv run python examples/bookstore/06_end_to_end.py

#	Example	What It Demonstrates
1	Schema Definition	Entities, fields, relationships, domains
2	Code Generation	Generate models, agents, GraphQL from schema
3	Data Agents	Deterministic CRUD, tool scoping, tracing
4	Domain Agents	LLM-powered orchestration and delegation
5	Auth & RBAC	Identity, JWT tokens, role-based permissions
6	End-to-End	Full pipeline: schema → agents → auth → query

Optional: Enable LLM features

Data agents, code generation, and RBAC work without an API key. For LLM-powered features (domain agents, conversational setup):

export GOOGLE_API_KEY="your-gemini-api-key"

Project Structure

NinjaStack is a modular monorepo of 15 focused packages:

ninja-stack/
├── libs/                          # Reusable libraries
│   ├── ninja-core/                # ASD schema models (entity, domain, relationship)
│   ├── ninja-agents/              # ADK agents (DataAgent, DomainAgent, Coordinator)
│   ├── ninja-auth/                # Auth gateway, strategies, RBAC
│   ├── ninja-codegen/             # Jinja2 code generation engine
│   ├── ninja-introspect/          # Database schema discovery
│   ├── ninja-persistence/         # Unified polyglot persistence
│   ├── ninja-gql/                 # Strawberry GraphQL generation
│   ├── ninja-boundary/            # Data tolerance & coercion
│   ├── ninja-graph/               # Graph-RAG bootstrapper
│   ├── ninja-models/              # Pydantic model generation
│   ├── ninja-deploy/              # K8s/Helm deployment pipeline
│   ├── ninja-ui/                  # CRUD viewer & chat UI generation
│   └── ninja-cli/                 # CLI tooling
├── apps/                          # Deployable applications
│   ├── ninja-api/                 # FastAPI server
│   └── ninja-setup-assistant/     # Gemini-powered conversational setup
├── examples/                      # Bookstore walkthrough (6 examples)
├── docs/                          # MkDocs source
└── site/                          # Landing page + built docs

Tech Stack

Layer	Technology
Language	Python 3.12+ · Pydantic v2
Agents	Google ADK · LiteLLM (model-agnostic)
API	FastAPI · Strawberry GraphQL
Auth	JWT · OAuth2 · API Keys · bcrypt
Persistence	SQLAlchemy · Motor/Beanie · Neo4j · ChromaDB
Deploy	Kubernetes · Helm
Package Mgmt	uv

Contributing

Setup

git clone https://github.com/codeninja/ninja-stack.git
cd ninja-stack
uv sync

Run tests

# Full suite
uv run pytest

# Specific library
uv run pytest libs/ninja-core/
uv run pytest libs/ninja-agents/
uv run pytest libs/ninja-auth/

# With coverage
uv run pytest --cov

Project conventions

• Commits: Conventional Commits — feat(ninja-agents): add tool scoping
• Branches: feat/issue-<N>-description from main
• PRs: One feature per PR, linked to an issue
• Tests: Every library has its own test suite. All tests must pass before merge.
• Code generation: Templates live in libs/ninja-codegen/src/ninja_codegen/templates/
• Adding a library: Create under libs/, add to root pyproject.toml workspace members

Build docs locally

uv run mkdocs serve
# → http://localhost:8000

Links

• 🏠 Homepage: codeninja.github.io/ninja-stack
• 📚 Documentation: codeninja.github.io/ninja-stack/docs
• 📖 Examples: examples/bookstore/
• 🐛 Issues: github.com/codeninja/ninja-stack/issues

License

TBD