# GPT Pilot — AI Developer That Builds Apps from Scratch

> Autonomous coding agent that writes full applications step-by-step, asking for human input at decision points.

## Install

Save in your project root:

# GPT Pilot — AI Developer That Builds Apps from Scratch

## Quick Use
```bash
git clone https://github.com/Pythagora-io/gpt-pilot && cd gpt-pilot
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python main.py
```

## Introduction
GPT Pilot is an autonomous coding agent that takes a project description and builds a working application from scratch. It follows a structured development process — planning, coding, testing, and debugging — and asks the developer for input at key decision points rather than working entirely unattended.

## What GPT Pilot Does
- Takes a natural-language app description and generates a full development plan
- Writes application code file by file, following the plan step by step
- Runs the application and iterates on bugs through automated debugging cycles
- Asks the developer for clarification or approval at important decision points
- Tracks project state so development can be paused and resumed

## Architecture Overview
GPT Pilot orchestrates multiple LLM calls through a state machine that models the software development lifecycle. Each phase (specification, planning, coding, testing, debugging) has dedicated prompts and parsing logic. The system maintains a conversation history and project context that carries across phases. It shells out to run code and captures output for automated debugging.

## Self-Hosting & Configuration
- Clone the repository and install Python dependencies in a virtual environment
- Set OPENAI_API_KEY or configure alternative LLM providers in the .env file
- Optionally configure a PostgreSQL database for persistent project storage
- Adjust token limits and model selection in the configuration file
- Run via CLI or integrate with the Pythagora VS Code extension

## Key Features
- Human-in-the-loop design that asks for developer input at critical decisions
- Structured development phases from specification through debugging
- Automatic bug detection and self-healing through iterative test-run cycles
- Project state persistence for pause and resume across sessions
- Support for multiple LLM backends including GPT-4 and Claude

## Comparison with Similar Tools
- **MetaGPT** — uses multiple specialized agent roles; GPT Pilot follows a linear dev workflow with human checkpoints
- **Aider** — pair programming on existing code; GPT Pilot builds full apps from scratch
- **OpenHands** — browser-based agent environment; GPT Pilot runs as a CLI tool
- **GPT Engineer** — generates initial codebases; GPT Pilot also handles iterative debugging
- **Cline** — IDE-integrated coding agent; GPT Pilot operates as a standalone process

## FAQ
**Q: What types of apps can GPT Pilot build?**
A: It works best with web applications (Python/Node.js backends, React frontends) but can handle CLI tools and scripts too.

**Q: Does it require GPT-4?**
A: GPT-4 or Claude produce the best results. GPT-3.5 works but with more errors and retries.

**Q: Can I resume a partially built project?**
A: Yes. GPT Pilot saves project state and can continue from the last checkpoint.

**Q: How does it handle bugs in generated code?**
A: It runs the code, captures errors, and feeds them back to the LLM for automated fix attempts.

## Sources
- https://github.com/Pythagora-io/gpt-pilot
- https://www.pythagora.ai

---
Source: https://tokrepo.com/en/workflows/6a4c20b8-3e8b-11f1-9bc6-00163e2b0d79
Author: AI Open Source