GPT Engineer

GPT Engineer's core bet is that the spec — not the editor — is where a project should begin. You write requirements into a prompt file, run gpte against an empty project directory, and get source files written into the workspace. The clarification loop is the part that matters in practice: before generating, the tool asks questions about what you actually want, and you can run the improvement option (-i) against an existing codebase to propose and apply changes. That two-mode workflow (new project, improve existing) is simple to hold in your head, which is more than most scaffolders manage. What you get is a real project structure: backend logic, frontend components, config files, plus npm and pip dependency management so the scaffold isn't just loose files. Python 3.10-3.12 is supported, with legacy 3.8/3.9 up to release 0.2.6. You can run it locally in Docker or in the browser, and you bring the model — open-source options like WizardCoder, an OpenAI-compatible local server such as llama.cpp (set OPENAI_API_BASE, OPENAI_API_KEY, MODEL_NAME and LOCAL_MODEL), or Azure-hosted endpoints. A customizable AI agent identity sits on top, and there's a bench interface for comparing custom agent implementations on supported public coding datasets. The honest weaknesses. Output quality is the model's, so a weak local endpoint produces weak scaffolds — this is not a tool that compensates for the model behind it. It's web-app-focused, so mobile, data pipelines and game engines are out. Complex state management or real-time sync will exceed what a spec-and-generate pass gives you, and you'll be writing the detailed, structured spec yourself. The browser option lowers the entry barrier, but the primary workflow is still a terminal and a Python environment. And the usual AI-scaffold caveat applies: complex instructions can be misread, so you review generated code before it ships. Where it fits: solo developers and founders who want starter code fast without a subscription, teams that need an MIT-licensed CLI they can self-host and modify, and educators demonstrating how natural language maps to executable code. Where it doesn't: non-technical teams wanting a visual canvas, and anyone who needs vendor support SLAs or enterprise compliance review around their build tooling.