The LLM Maze Agent framework provides a Python-based environment for building intelligent agents capable of navigating grid mazes using large language models. By combining modular environment interfaces with chain-of-thought prompt templates and heuristic planning, the agent iteratively queries an LLM to decide movement directions, adapts to obstacles, and updates its internal state representation. Out-of-the-box support for OpenAI and Hugging Face models allows seamless integration, while configurable maze generation and step-by-step debugging enable experimentation with different strategies. Researchers can adjust reward functions, define custom observation spaces, and visualize agent paths to analyze reasoning processes. This design makes LLM Maze Agent a versatile tool for evaluating LLM-driven planning, teaching AI concepts, and benchmarking model performance on spatial reasoning tasks.