gpt-engineer: AI Code Generation CLI
gpt-engineer is a Python CLI that turns a plain-text prompt file into a generated, runnable codebase, then lets you iterate on that code with a follow-up prompt and the `-i` flag. Reach for it if you want a scriptable, terminal-first way to spin up a small prototype from a written spec. Skip it if you want an actively evolving, hackable agent — the README itself sends that use case to aider.
What is gpt-engineer
gpt-engineer is a command-line tool that generates an entire project from a natural-language prompt file, then executes the resulting code for you. You write instructions into a file literally named `prompt` inside a project folder and run `gpte <project_dir>`; the AI writes and runs the code, and a follow-up `-i` run can revise it. It defaults to OpenAI and Anthropic models, with documented paths to Azure OpenAI and local models.
Key features of gpt-engineer
- ✓Generates a full project from a single `prompt` text file inside an empty folder — no chat window involved.
- ✓Improve mode (`gpte <project_dir> -i`) reads your existing code and applies changes described in an updated prompt.
- ✓Preprompts folder overrides the agent's instructions via `--use-custom-preprompts`, so you can make it remember conventions across projects.
- ✓Vision input: pass `--image_directory` with a vision-capable model to feed UX or architecture diagrams in as context.
- ✓Ships a separate `bench` binary for benchmarking your own agent implementations against the APPS and MBPP datasets.
- ✓Supports OpenAI, Azure OpenAI, and Anthropic models by default, with a documented path to local models like WizardCoder.
- ✓Docker instructions and a dedicated Windows README ship alongside the CLI.
Installing gpt-engineer
For a stable release: `python -m pip install gpt-engineer`. For development: clone the repo, then run `poetry install` followed by `poetry shell` to enter the virtual environment. Python 3.10-3.12 is actively supported; version 0.2.6 was the last release to support Python 3.8-3.9. Next, set an API key — export `OPENAI_API_KEY`, or copy `.env.template` to `.env` and add it there — or wire up a custom/local model per the linked docs. A Docker path and a separate Windows README are also provided.
Generating and improving code with gpt-engineer
Create an empty folder, add a file named exactly `prompt` (no extension) with your instructions, then run `gpte <project_dir>` — for example `gpte projects/my-new-project`. gpt-engineer writes and executes the code. To improve it later, edit the `prompt` file with your requested changes and rerun with `-i`, e.g. `gpte projects/my-old-project -i`. For vision input, add `--image_directory` and a vision-capable model, e.g. `gpte projects/example-vision gpt-4-vision-preview --prompt_file prompt/text --image_directory prompt/images -i`.
Strengths
- ✓The whole interaction model is a text file and one CLI command, which makes it easy to script or wrap in other tooling.
- ✓Ships its own `bench` binary for benchmarking custom agents against public datasets (APPS, MBPP), not just for running the CLI.
- ✓Preprompts are plain files you can override, so tuning agent behavior doesn't require touching the Python source.
- ✓Works with OpenAI, Azure OpenAI, and Anthropic out of the box, with a documented route to local models.
Limitations of gpt-engineer
- △There's no chat interface or IDE plugin — every iteration means editing a `prompt` file and rerunning a CLI command.
- △The README frames gpt-engineer as the less actively developed option, pointing to aider for a well-maintained, hackable CLI instead.
- △Local/open-source model support (e.g. WizardCoder) needs extra setup beyond the default OpenAI/Anthropic configuration, per the docs.
- △Vision input only kicks in when you pass a vision-capable model explicitly — it isn't the default mode.
Alternatives to gpt-engineer
Frequently asked questions
gpt-engineer itself is free and open source under the MIT license. You still need an API key from a model provider (OpenAI, Azure OpenAI, or Anthropic by default), and using that provider's API carries its own cost.
gpt-engineer actively supports Python 3.10 through 3.12. Python 3.8 and 3.9 are only supported by the older 0.2.6 release.
gpt-engineer can run local or open-source models such as WizardCoder instead of OpenAI, though the README notes this needs extra setup beyond the default configuration — see the linked docs for the steps.
gpt-engineer's own README points to aider as the better-maintained, hackable CLI option, and positions itself as the original experimentation platform the space grew out of. For ongoing active development, the README's own advice is to look at aider.
gpt-engineer supports image input for vision-capable models via the `--image_directory` flag, letting you pass UX or architecture diagrams as extra context alongside your text prompt.
The `bench` binary that ships with gpt-engineer lets you benchmark your own custom agent implementations against public coding datasets, currently APPS and MBPP.
The problem it solves
Before gpt-engineer, turning a short spec into working code meant hand-writing it or manually copy-pasting a chat model's replies into an editor. gpt-engineer's own description calls it the original code-generation experimentation platform — the precursor to lovable.dev — and it's built specifically around a `prompt` file and one CLI command so an AI can write and execute a project's code directly, with no human relaying snippets in between.
Best use cases
- •Turning a short written spec into a runnable prototype without a back-and-forth chat session.
- •Feeding a UI mockup or architecture diagram to a vision-capable model to scaffold a matching project.
- •Benchmarking a custom coding-agent implementation against the APPS or MBPP datasets via `bench`.
- •Iterating on an existing folder of code by rewriting the `prompt` file and rerunning with `-i`.
Who should try it — and who should skip
Try gpt-engineer if you're comfortable in a terminal and want to turn a written spec into a runnable prototype without a chat UI, or if you're building and testing your own coding agent and want the `bench` harness for APPS/MBPP. Skip it if you want a tool under active headline development — the project's own README redirects that need to aider — or if you'd rather have a managed, git-controlled UI, which is what gptengineer.app provides.
Related repositories
Is gpt-engineer worth your time?
ChatGPT, Claude and Perplexity can all read this page. Ask one of them what it makes of gpt-engineer.
