Stirrup is a lightweight framework designed for creating agents that interact with various models in a flexible manner. It prioritizes allowing the model to dictate its approach rather than forcing it into a predefined workflow. This flexibility aims to enhance performance and results, contrasting with many existing frameworks that can hinder agent capabilities. The framework includes built-in best practices derived from analyzing leading agents like Claude Code and Codex, particularly in areas like context management and foundational tools. The Python implementation of Stirrup offers features that enable local code execution, web browsing, and connections to MCP servers. Users can customize their agents extensively, either using Stirrup as a package or as a base template. Installation is straightforward, with options to include specific components based on needs. For example, running `pip install 'stirrup[all]'` installs all optional tools, while individual extras can be added as needed. Stirrup supports a range of functionalities, such as a human-in-the-loop feature that allows user input during agent execution. It also includes a skills system for modular instruction packages and manages context automatically to prevent overload. The framework supports multimodal processing, enabling the handling of images, video, and audio. Developers can create agents using predefined tools or build their own, making it suitable for both simple and complex applications. The documentation provides thorough guidance on installation, core concepts, and creating custom tools.