// the find
AntonOsika/gpt-engineer
CLI platform to experiment with codegen. Precursor to: https://lovable.dev
gpt-engineer is one of the original LLM codegen CLIs: point it at a prompt file and it writes a whole project from scratch, or takes an -i flag to iterate on existing code. It's the open-source precursor to the hosted gptengineer.app/Lovable product, so today it's mostly historically interesting rather than actively pushed forward.
Clean separation between agent logic and execution environment (core/base_agent.py, base_execution_env.py), which makes it easy to swap in a different model or sandbox. The benchmark module (APPS, MBPP) is a real, runnable harness for comparing agent implementations, not just a marketing claim. preprompts/ as plain editable text files is a simple, hackable way to change agent behavior without touching code.
Last push was May 2025 and the README itself points you to aider for anything you'd actually want maintained — that's the project admitting it's no longer the actively developed option. Whole-project generation from a single prompt file doesn't hold up on anything beyond toy projects; context window and lack of incremental review make it unsuitable for real codebases. OpenAI/Anthropic-only by default, with local model support bolted on via a docs page rather than first-class. No CI coverage for the diff-application logic (core/diff.py, chat_to_files.py) which is exactly the part that silently corrupts files when the model's output format drifts.