A lightweight, customizable framework for building AI agents, Stirrup empowers developers by harmonizing with LLMs rather than imposing rigid workflows. It integrates best practices for context management and foundational tools, offering a flexible starting point for creating sophisticated, domain-specific agents.

How It Works

Stirrup's core philosophy is to "get out of the way" and let the LLM dictate the task execution strategy, mirroring successful approaches like Claude Code. This contrasts with frameworks that enforce strict, potentially performance-degrading, workflows. It incorporates analyzed best practices for context management and foundational tools, providing a robust yet adaptable architecture. Developers can use it as a package or clone it as a template for deep customization.

Quick Start & Requirements

Installation is straightforward via pip: pip install stirrup for the core, or pip install 'stirrup[all]' for optional components. Key dependencies include environment variables for API keys, such as OPENROUTER_API_KEY for the example client and BRAVE_API_KEY for web search functionality. The project supports Python and can be managed with tools like uv. Full documentation is available at artificialanalysis.github.io/Stirrup.

Highlighted Details

• Built-in Tools: Includes essential utilities like online search, web fetching, and code execution (supporting local, Docker, and E2B sandbox environments).
• Extensibility: Features a "Skills system" for modular, domain-specific extensions and a generic Tool class for easy custom tool definition.
• Provider Flexibility: Offers pre-built support for OpenAI-compatible APIs and LiteLLM, allowing integration with various LLM providers (e.g., Anthropic, Google) and custom clients.
• Advanced Features: Incorporates automatic conversation history summarization for context management and supports multimodal inputs (image, video, audio) with format conversion.

Maintenance & Community

The provided README does not detail specific contributors, sponsorships, or community channels like Discord or Slack.

Licensing & Compatibility

The project is licensed under the permissive MIT LICENSE, generally allowing for broad compatibility with commercial and closed-source applications without copyleft restrictions.

Limitations & Caveats

Web search functionality is contingent on the presence of a BRAVE_API_KEY. While flexible, integrating non-default LLM providers may require specific client configurations or the installation of optional dependencies like stirrup[litellm].