SWE-agent: Use Language Models to Fix GitHub Issues

Overview

SWE-agent is a Python research tool that connects a language model to tools for working in software repositories. Given a GitHub issue, it can attempt to understand the problem, inspect and edit the code, and produce a proposed fix.

It is aimed at researchers and developers exploring language-model-driven software engineering, rather than being a general-purpose autonomous coding service. The repository also documents uses for custom coding tasks and, through its EnIGMA mode, cybersecurity challenges. The maintainers say mini-SWE-agent has superseded this project and recommend it for new use.

Key Features

• Accepts a GitHub issue as the starting point for an automated software-fixing attempt.
• Lets users choose a language model, with the README naming GPT-4o and Claude Sonnet 4 as examples.
• Provides tools for a model to inspect and modify real code repositories.
• Supports configuration through a YAML file.
• Includes workflows for benchmarking on SWE-bench.
• Offers a documented EnIGMA mode for capture-the-flag cybersecurity challenges, with a note to use version 0.7 while EnIGMA is updated for 1.0.
• Is designed to be hackable for research and custom tasks.

Use Cases

• Software engineering researchers can study how language models interact with repository tools and attempt issue fixes.
• Developers evaluating coding agents can run the project against repository issues or SWE-bench tasks.
• Teams with custom coding challenges can adapt the configurable agent workflow to tasks beyond issue resolution.
• Cybersecurity researchers and CTF participants can explore EnIGMA, using the recommended 0.7 version while its 1.0 update is pending.

Project Facts

• Language: Python
• License: MIT
• Stars: 20.5k
• Forks: 2.2k
• Topics: agent, agent-based-model, ai, cybersecurity, developer-tools, llm, lms
• Archived: No

Getting Started

The README links to installation instructions and a command-line hello-world guide. Start with the official documentation, then follow the hello-world guide. The README also describes a browser-based GitHub Codespaces option. Consult the repository README for current setup details.

Alternatives

• DeepSeek-Reasonix: DeepSeek-Reasonix is a persistent terminal coding agent for ongoing development, while SWE-agent focuses on configurable software-engineering tasks such as resolving GitHub issues.
• pi: Pi is a general-purpose, extensible agent toolkit with a coding CLI, while SWE-agent is purpose-built for software-engineering tasks in repositories.
• Ouroboros: Ouroboros emphasizes autonomous development, persistent memory, and self-modification, while SWE-agent targets issue-solving workflows and software-engineering research.
• aider: Aider is an interactive terminal assistant for making repository changes with a chosen model, while SWE-agent is designed to tackle software-engineering tasks autonomously.

Considerations

• The project is not the maintainers' recommended choice for new deployments: development effort has shifted to mini-SWE-agent, which they describe as simpler and recommend going forward.
• Results depend on the selected language model and the agent's configuration; the project does not remove the need to review generated changes.
• EnIGMA guidance specifically points users to SWE-agent 0.7 until its compatibility with version 1.0 is updated.
• The project is research-oriented and expects users to consult its documentation for setup, model configuration, and benchmark workflows.