The CodeAct approach consolidates LLM agent actions into a unified code-based action space, enabling agents to execute code, observe results, and dynamically revise or emit new actions through multi-turn interactions. The repository releases the CodeActAgent model (fine-tuned Mistral-7B), a multi-turn instruction dataset (CodeActInstruct), and a benchmark (M³ToolEval). It supports serving via Ollama, llama.cpp for local laptop inference, and Kubernetes deployment.