AI Engine Framework

The distribution is published to PyPI as sumbai. There is no CI badge because this project has no public source repository — see Documentation.

A modular, extensible framework for building AI-powered applications with Python.

Features

• Modular Architecture - 17 core modules for agents, memory, tools, workflows, and more
• Plugin System - Contracts that let capabilities be added without touching core. Note: the working providers (OpenAI-compatible, Anthropic, Gemini) are implemented in-core; the plugins/ directory in the source tree is not part of the published wheel and there are no separate provider packages to install
• Developer SDK - Fluent API with async support and type safety
• CLI Interface - Interactive chat, project management, and diagnostics
• Observability - Built-in tracing, metrics, and audit logging
• Multi-Agent - Collaboration, organization, and session management
• Production Ready - Comprehensive testing, documentation, and CI/CD

Installation

pip install sumbai

The import names stay ai_engine, sdk, and cli; the ai command-line entry point is installed with the base package.

With optional extras:

pip install "sumbai[embeddings,postgres,otel,pdf,docx,xlsx,litellm]"

The package declares nine extras in total — the seven above plus dev (test and lint toolchain) and interactive (InquirerPy, already a hard dependency and kept only for compatibility). See Installation for what each one brings.

There is no openai, anthropic, qdrant, redis, or all extra. Model providers are implemented in-core over httpx and need no extra package. Asking for an extra that does not exist is not an error, and pip will not tell you: measured with pip 26.1.2, pip install "sumbai[qdrant,openai]" prints Successfully installed sumbai-1.0.1 and nothing else, so a typo here fails completely silently. (uv pip install does warn; pip does not.)

Quick Start

Verify the install:

python -c "from sdk import AISDK; print('SDK installed successfully')"
ai --version

A fresh AISDK() has no model provider wired, so asking it to answer something fails honestly rather than inventing a reply — sdk.agents.run(...) raises No LLM provider configured for ChatLoop, and the older Engine().run("Hello, world!") returns a Response carrying CAPABILITY_NOT_FOUND: Capability 'chat' not found. Both are measured, and neither is a bug: configuring a provider is the caller's step.

Wiring one is a DI override today, not a facade call. The worked example is samples/sdk_tour/05_agent.py, and samples/quickstart/chat.py is a complete runnable chat client.

Documentation

Documentation lives in this repository; there is no published documentation site. docs/ and samples/ are not part of the distribution — the sdist carries ai_engine, engine, sdk, cli and these top-level files, so the links below resolve in a clone of the source and not in an installed package.

• Getting Started
• Core Concepts
• SDK Reference
• CLI Reference
• Cookbook
• Changelog

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Security

For security vulnerabilities, please see SECURITY.md.

License

This project is licensed under the MIT License - see the LICENSE file for details.