wood-q/mokioagent

MokioClaw is a teaching project written in Python that guides you through building an AI agent system from scratch, step by step. Rather than shipping a finished all-purpose assistant, it is designed as a learning path, starting from a minimal tool-calling example and growing it into something capable of editing code, running a terminal, and reviewing its own work. The central showcase task is called Mini Game Studio: you give the agent a natural-language instruction (such as "create a simple Python guessing game"), and it follows a planner-executor-reviewer loop to write the code, run basic checks, and report what it did. The agent keeps two context files (a notepad for cross-task facts and a to-do list for the current task) that are injected into every AI call so the model always knows where it left off. File operations and shell commands are sandboxed to a workspace folder, and only Python and pytest commands are permitted, blocking chained shell tricks for safety. You would use this if you are learning how AI agents work under the hood and want a clear, iterative example rather than a black-box framework. The project is written in Python and uses LangGraph for the plan-and-execute flow, with a command-line interface built on Typer. It connects to any OpenAI-compatible API endpoint, so you can swap in different language models.