AntonOsika/gpt-engineer

Specify what you want it to build, the AI asks for clarification, and then builds it.

Summary

gpt-engineer is a Python CLI for code-generation experiments that turns natural-language instructions in a prompt file into a working project. It supports OpenAI and Anthropic models by default, with optional vision input, custom pre-prompts, and a benchmarking tool. Note: the repository is marked as archived, so it is best treated as a reference platform rather than an actively maintained tool.

What it is useful for

The project targets engineers who are building or studying autonomous coding agents. Documented use cases include:

• Bootstrapping a new codebase from a written specification in a prompt file
• Iterating on existing code by running with the -i flag and describing desired changes
• Feeding image inputs (UX or architecture diagrams) alongside text prompts when using a vision-capable model
• Customizing the agent's "identity" by overriding the preprompts folder via the --use-custom-preprompts argument
• Benchmarking custom agent implementations against APPS and MBPP using the bench binary

It also exposes hooks for alternative backends (Azure OpenAI, local models such as WizardCoder), though the README directs users to the documentation for setup details. For users who want a managed web UI, the project points to gptengineer.app, and for a hackable CLI alternative, it points to aider.

How engineers can use it

The README documents two install paths. For a stable install: python -m pip install gpt-engineer. For development: clone the repo, then run poetry install and poetry shell.

Next, configure an OpenAI key by either exporting OPENAI_API_KEY as an environment variable, or copying .env.template to .env and adding the key there. Windows users are pointed to a separate WINDOWS_README.md; a Docker path is documented under docker/README.md, and GitHub Codespaces is offered for browser-only use.

The basic workflow is:

1. Create an empty project directory anywhere on disk.
2. Add a prompt file (no extension) containing your instructions.
3. Run gpte <project_dir> to generate code, or gpte <project_dir> -i to improve code that already lives in that directory.

For example, gpte projects/my-new-project from the repo root generates a new app, while gpte projects/example-vision gpt-4-vision-preview --prompt_file prompt/text --image_directory prompt/images -i demonstrates the vision flow.

Documented limitations to keep in mind: the repository is archived, only Python 3.10–3.12 is actively supported (the last release for 3.8–3.9 was 0.2.6), and vision or non-OpenAI model usage requires extra setup covered in the linked docs rather than the main README. Running the tool also implies agreeing to the project's terms of use.