About GitHub
OpenAGI is an open-source Python package for creating AI agents that integrate large language models (LLMs) with domain-specific tools and experts. It provides a structured framework for developing, uploading, and sharing agents, and is designed to work with the AIOS platform. The project is based on research published at NeurIPS 2023 (paper: 'OpenAGI: When LLM Meets Domain Experts') and is available under the MIT license. Installation is simple via pip, and users can extend the system by adding custom agents with configuration and dependency files.
Key Features
Agent creation framework with structured templates
Integration with AIOS platform
Support for external tools and custom tool development
Agent upload and download for sharing
Based on academic research published at NeurIPS 2023
Extensible via custom agents with config and dependencies
Pros & Cons
Pros
- Open source under MIT license
- Easy installation via pip (pip install pyopenagi)
- Well-structured agent template for quick development
- Backed by academic research at NeurIPS 2023
- Supports community contributions via upload/download
Cons
- Documentation appears limited to basic setup and usage
- Requires familiarity with the AIOS ecosystem
- Relatively new project with a smaller community (2.3k stars)
Best For
Building domain-specific AI agents for various tasksResearch and experimentation with LLM-based agent systemsPrototyping agent workflows and tool integrationSharing and reusing agent implementations within the community
FAQ
What is OpenAGI?
OpenAGI is an open-source package for creating AI agents that integrate large language models with domain experts. It is used to build agents for the AIOS platform.
How do I install OpenAGI?
You can install via pip: pip install pyopenagi, or clone the repository and run pip install -e . for local development.
How can I add a new agent?
Create a folder under pyopenagi/agents with the author and agent name, including agent.py, config.json, and meta_requirements.txt.
How to upload or download agents?
Use the interact.py script with --mode upload or --mode download and specify the agent path (e.g., author_name/agent_name).