SWE-agent
Autonomous LM agent that uses tools to fix real GitHub repository issues.
Empower your language model of choice to resolve real repository issues and cybersecurity challenges.
SWE-agent allows language models like GPT-4o or Claude to autonomously use tools and fix issues in real GitHub repositories. Developed by researchers from Princeton and Stanford, it provides a free-flowing and generalizable framework for developers and researchers.
What it is
The tool takes a GitHub issue and tries to automatically resolve it using a language model you configure. It operates via a single YAML configuration file and gives maximal agency to the LM to run commands and inspect code.
Who it's for
It suits AI researchers and developers looking to automate software engineering tasks or benchmark agent performance. It does not suit casual users who want a simple point-and-click interface without managing an LM backend.
Key features
Autonomous Issue Fixing
Resolves issues in real GitHub repositories using your language model of choice.
YAML Configuration
Behavior is governed entirely by a single easily modifiable YAML file.
Offensive Cybersecurity
Includes EnIGMA mode for solving capture-the-flag and security challenges.
Research-Oriented Design
Simple and hackable architecture built by Princeton and Stanford researchers.
Up and running in 3 steps
Our honest verdict
What's great
- Achieves state-of-the-art performance on SWE-bench among open-source projects.
- Provides maximal agency to the language model for free-flowing task execution.
- Fully configurable and documented using simple YAML files.
- Includes specialized modes for offensive cybersecurity and competitive coding.
Where it falls short
- Requires ongoing API costs for the underlying language model you choose to run it with.
- Active development has largely shifted to mini-swe-agent, which is simpler and recommended going forward.
- Setting up and benchmarking requires familiarity with command-line tools and Python environments.
Pricing & license
Free under its MIT license, but you will incur usage costs from whichever commercial language model API you choose to connect to it.
Open remote AI-training roles
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
Which language models are supported?
You can use the language model of your choice, such as GPT-4o or Claude Sonnet.
Should I use SWE-agent or mini-swe-agent?
The maintainers recommend using mini-swe-agent going forward as it matches SWE-agent performance while being much simpler.
Can it be used for things other than GitHub issues?
Yes, it can also be employed for offensive cybersecurity capture-the-flag challenges or custom coding tasks.
