GPT Engineer

Type: full-code · Vendor: Anton Osika / community · Language: Python · License: MIT · Status: abandoned · Status in practice: deprecated · First released: 2023-04-30

Links: homepage repo

Early spec-first code-generation CLI: the user writes a natural-language prompt file describing the software, gpte generates the project end-to-end, and -i mode iterates on improvements.

Description. GPT Engineer was one of the original LLM coding agents — a Python CLI that turned a `prompt` file into a generated codebase, with an improvement loop via `gpte <project> -i`. The repo was archived by the owner on 2026-04-22 and explicitly redirects users to aider or to its commercial successor Lovable. Historically significant as the prototypical spec-first single-shot generator: drop in a prompt, watch the AI write and execute the project, then iterate.

Agent loop shape. Spec-first one-shot generation followed by optional improvement passes. Phase 1: read the prompt file, generate the project, execute. Phase 2 (-i): take an existing codebase, ask the model for improvements, apply, repeat. Minimal tool-use surface compared to modern agents — the loop is generate → write files → execute, not a multi-tool ReAct loop.

Primary use cases

single-shot project generation from a prompt file
iterative improvement of a generated project via -i flag
experimenting with pre-prompts and custom agents via the benchmark harness

flowchart TD prompt[Prompt file] --> gpte[gpte CLI] gpte --> gen[LLM generates project] gen --> write[Write files] write --> exec[Execute project] exec --> done[Done] done -.->|-i flag| improve[Ask AI to improve] improve --> gen

Key concepts

Spec file (prompt) → spec-first-agent (docs) — Plain-text natural-language specification of the desired software.
Improve mode (-i) → spec-driven-loop — Ask the AI to improve an existing project.
Pre Prompts — Customise agent behaviour via --use-custom-preprompts.
bench — Benchmark harness for evaluating custom agents.