GPT Engineer

What is GPT Engineer?

GPT Engineer is an open-source command-line program that writes software from a plain-English brief. You describe the app in a file called prompt, run one command, and a language model generates the project files. A second mode edits existing code to match new instructions.

Anton Osika started it in 2023, and it passed 55,000 GitHub stars. Its hosted spin-off, gptengineer.app, became Lovable in December 2024. The last PyPI release, 0.3.1, shipped in June 2024, and the repository was archived on 22 April 2026.

Key Features

• Prompt-to-codebase: GPT Engineer builds a multi-file project from one plain-English specification in a file called prompt. A single gpte command starts the run, and the generated files land in your project folder.
• Clarification step: The agent can ask clarifying questions before it writes any code. Answering them gives the model more context, which reviewers say improves the first version of the project.
• Improve mode: The -i flag applies requested changes to an existing codebase. You describe the change in the prompt file, and the agent edits the code rather than starting again.
• Model choice: GPT Engineer works with OpenAI, Azure OpenAI and Anthropic models. Open models such as WizardCoder can also connect through compatible local servers with extra configuration.
• Custom preprompts: You can override the preprompts folder that defines the agent’s identity. This lets you change coding conventions and keep knowledge between projects. Teams can tune the agent for a house style without editing the core code.
• Vision input: The –image_directory flag adds diagrams or images as context for vision-capable models. UX sketches and architecture diagrams can then guide the generated structure, giving the model more than a written description to work from.
• Benchmark command: A bench tool tests custom agents on the APPS and MBPP datasets. Researchers use it to compare their own coding agents against a shared baseline.
• Local output: Generated code is written as plain files to a local folder. You can inspect every file, commit it to Git and run it with your usual tools.

Pros

• No licence fee: The MIT licence allows free use, modification and forking. Your only running cost is the model API you connect, billed by your provider. There is no subscription and no vendor lock-in to the project itself.
• Transparent design: Every prompt and step in the agent is readable. That makes GPT Engineer a clear teaching example of how early coding agents planned and wrote software.
• Local control: Files stay on your machine and requests use your own API keys. You avoid handing code to a hosted platform, which suits privacy-conscious experiments and code you would rather not upload.
• Multi-provider flexibility: Switching between OpenAI, Azure, Anthropic and local models avoids lock-in to one vendor. You can test a cheaper model first and upgrade only when needed.
• Hackable base: Custom preprompts and the benchmark tool make it practical to build and test your own agent. The project passed 55,000 GitHub stars, so many forks and examples exist.
• Historic value: GPT Engineer shows how Lovable began. Its hosted spin-off, gptengineer.app, became Lovable in December 2024, which makes the code a useful record of that product’s origins.

Cons

• Archived project: The repository was archived on 22 April 2026 and is now read-only. Issues and pull requests will not be answered, and no further fixes will arrive.
• Ageing defaults: Prompts were tuned for 2023 and 2024 models, and the last PyPI release, 0.3.1, shipped in June 2024. Newer models may behave differently without prompt changes.
• Technical setup: You need Python 3.10 to 3.12, a terminal and your own API key. Non-developers will find hosted app builders far easier to start with, since GPT Engineer offers no graphical installer or web interface.
• CLI only: GPT Engineer has no web interface or visual editor. Every change happens through prompt files and terminal commands, which slows down design-heavy work. Frontend layouts are hard to judge until you run the generated app yourself.
• Variable output: Generated projects often need manual debugging, especially on complex apps. Treat results as prototypes rather than production-ready code, and budget time to fix errors before shipping anything.
• Single-user tool: GPT Engineer has no collaboration features for teams. Shared work relies on Git and your own review process outside the tool. Each developer runs the agent separately with their own API key and settings.

Who is Using GPT Engineer?

• Researchers: Teams use it as a baseline coding agent.
• Python developers: Hobbyists scaffold quick prototypes from the terminal.
• Educators: Teachers use it to show how coding agents work.
• Forkers: Developers build custom agents on the MIT code.

Pricing

The tool is free, and your only spend is the model API you connect (checked October 2026).

Lovable is a separate product with its own plans. API rates shift often, so check your provider before large runs.

What Makes GPT Engineer Unique?

It was among the first public tools to generate a whole repository from one prompt. Its fast GitHub growth proved demand for AI app builders and led its creators to found Lovable. Today it works best as a readable, hackable reference.

Compatibilities and Integrations

• Platforms: Runs on macOS, Linux and Windows with Python.
• Model providers: Supports OpenAI, Azure OpenAI and Anthropic APIs.
• Local models: Open models connect through compatible local servers.
• Output: Writes plain files to a local folder for Git.

GPT Engineer Tutorials

A first project takes only a few steps.

1. Run python -m pip install gpt-engineer.
2. Set your OpenAI or Anthropic API key.
3. Create a folder with a prompt file describing the app.
4. Run gpte with the folder path.
5. Use the -i flag later to improve the code.

How We Rated It

Summary

GPT Engineer suits developers and researchers who want to study an early prompt-to-code agent. Its main limitation is that the project is archived and unpatched. For a maintained terminal tool, the README points to Aider. Our tip is to test with a cheaper model, since each run makes several API calls.