RPG-ZeroRepo  by microsoft

LLM-driven framework for automated software repository generation

Created 3 months ago
446 stars

Top 67.3% on SourcePulse

GitHubView on GitHub
Project Summary

Summary

RPG-ZeroRepo automates software repository creation from natural language descriptions, targeting engineers and researchers. It employs a novel Repository Planning Graph (RPG) and an LLM-driven pipeline to translate high-level project goals into features, architecture, and fully implemented, tested source files, accelerating initial development.

How It Works

A three-phase LLM pipeline: 1) Property Level decomposes descriptions into feature trees and components (RPG-Encoder). 2) Implementation Level designs repository skeletons, interfaces, and tasks. 3) Code Generation executes tasks in Docker via the trae-agent, using an iterative TDD loop for code writing and validation. The RPG, a unified graph of functional and structural aspects, enables scalable, coherent codebase generation.

Quick Start & Requirements

Setup: Create my_project/checkpoints/ and my_project/workspace/, define goals in checkpoints/repo_data.json, configure LLM settings in configs/zerorepo_config.yaml and configs/trae_config.yaml. Run via python main.py --config ... --resume or bash scripts/run_main.sh. Prerequisites: Python, Docker, LLM API keys. Pipeline is resumable. Relevant papers: arXiv:2602.02084, arXiv:2509.16198.

Highlighted Details

  • Repository Planning Graph (RPG): Unified graph representing functional (features) and structural (files, dependencies) repository aspects.
  • End-to-End Generation: Automates codebase creation from natural language overview to version-controlled repository.
  • Iterative TDD Loop: Phase 3 uses a Dockerized TDD cycle for code generation and validation.
  • Resumable Pipeline: Supports checkpointing and resuming execution at any stage.

Maintenance & Community

Recent activity includes open-sourcing the codebase (Feb 12, 2026) and releasing research papers (arXiv:2602.02084, 2509.16198). RPG-Encoder release is pending. No community channels or contributor details are provided.

Licensing & Compatibility

The specific open-source license is not detailed in the README. Compatibility requires Python and Docker environments, with LLM provider integration being key.

Limitations & Caveats

RPG-Encoder is pending open-source release. A Claude Code plugin is under development. Functionality depends heavily on configured LLM providers and their API access. Code generation requires Docker.

Health Check
Last Commit

1 day ago

Responsiveness

Inactive

Pull Requests (30d)
7
Issues (30d)
5
Star History
448 stars in the last 30 days

Explore Similar Projects

Feedback? Help us improve.