GPT Engineer

Supports OpenAI, Azure, Anthropic, and open-source models like WizardCoder, allowing users to choose based on cost or performance needs. The documentation provides specific instructions for configuring local models.

Users can override the preprompts folder to define AI identity and memory across projects, enabling tailored coding styles. This is accessible via the --use-custom-preprompts argument.

Accepts image directories alongside text prompts for vision-capable models, useful for generating code from UX or architecture diagrams. This is specified with the --image_directory flag.

Includes a 'bench' binary for evaluating custom agent implementations against public datasets like APPS and MBPP, facilitating research and performance comparison as described in the benchmarking suite.