gpt-engineer

Describe software in natural language, watch an AI agent write, run, and improve it.

Free· Free and open source under MIT license. Users pay only for the underlying LLM API calls (OpenAI, Anthropic, Azure OpenAI) or run local models at zero token cost.AgentsOpenAI GPT (default), Anthropic Claude, Azure OpenAI, and open-weights models like WizardCoder via configuration

Visit website →

Best for

Developers, researchers, and tinkerers who want a hackable, model-agnostic CLI coding agent they can script, fork, or benchmark without paying for a platform.

Skip if

Teams needing an actively maintained, IDE-integrated agent for large production codebases — pick Aider, Cursor, Claude Code, or Cline instead.

gpt-engineer is one of the original open-source coding agents that helped kick off the 2023 wave of autonomous code generation. Anton Osika released it as a proof that a small CLI, a prompt file, and a capable LLM were enough to bootstrap entire projects from a paragraph of English. You drop a plain-text 'prompt' file into an empty directory, run `gpte <dir>`, and the agent decomposes the spec, asks clarifying questions when needed, writes source files, and can execute the resulting code. An `-i` (improve) flow lets it iterate on an existing codebase rather than starting fresh.

Under the hood it drives OpenAI GPT models by default, with support for Azure OpenAI, Anthropic, and self-hosted open-weights models such as WizardCoder via LiteLLM-style config. Vision-capable models can consume diagrams and screenshots as part of the spec. Custom preprompts let advanced users reshape the agent's planning, code-writing, and self-review steps without forking the repo. A benchmarking harness against APPS and MBPP is included, which makes it a reasonable starting scaffold for anyone researching coding-agent behavior.

The project's ceiling matches its era: it works best on greenfield scripts, small utilities, and single-file demos, and it struggles on large multi-module codebases where modern IDE-integrated agents (Cursor, Aider, Claude Code, Cline) have taken over. The repository was archived in April 2026, so it is now read-only and no longer receiving fixes. It remains genuinely useful as a hackable reference implementation, a teaching example of the plan-then-code loop, and a lightweight CLI for developers who want a scriptable, model-agnostic generator without a subscription.

Editor's take

A landmark project that showed what a plan-then-code loop could do with almost no scaffolding. Two years on, the archive banner tells the story: the ideas won, but the implementation was overtaken by IDE-native agents. Still worth cloning as a reference for how simple a coding agent can actually be.

— The AI Tool Bible editorial team

Pros

✅ Fully open source (MIT) with a small, readable codebase that's easy to fork or study.
✅ Model-agnostic: swap OpenAI, Azure OpenAI, Anthropic, or local open-weights models.
✅ Zero platform cost — you pay only for tokens, or nothing at all with local models.
✅ Simple CLI-first workflow (`gpte <dir>`) that scripts and CI can call.
✅ Built-in improve mode for iterating on existing code, not just greenfield generation.
✅ Preprompt customization lets you retune agent behavior without patching source.
✅ Ships with benchmarking against APPS and MBPP for coding-agent research.

Cons

⚠️ Repository was archived in April 2026 — no active maintenance, bug fixes, or new model support.
⚠️ Best suited to small greenfield projects; struggles on large multi-file codebases.
⚠️ No IDE integration — lives entirely in a terminal with a text prompt file.
⚠️ Requires bring-your-own API keys and manual configuration for non-OpenAI models.
⚠️ Python 3.10-3.12 only; older environments are unsupported.
⚠️ Newer agents (Aider, Cursor, Claude Code, Cline) have overtaken it on both quality and DX.

Use cases

Greenfield script and prototype generationSmall single-file utility creation from a specIterative code improvement via improve modeCoding-agent research and benchmarking (APPS, MBPP)Teaching example for LLM agent loopsScriptable code generation in CI pipelinesLocal-model code generation with self-hosted LLMs