SWE-agent

SWE-agent is the open-source autonomous coding agent that automatically fixes GitHub issues by reading the issue description, exploring the codebase, making code changes, and submitting pull requests using whatever LLM the user configures. Developed by researchers at Princeton University and Stanford University and published as a NeurIPS 2024 paper, SWE-agent has become one of the most academically-credentialed autonomous coding agents and achieves state-of-the-art results on the SWE-bench leaderboard among open-source projects. The GitHub repository has 19,800+ stars and 2,200+ forks. Important: the SWE-agent team now recommends mini-swe-agent as the primary tool for new users. Mini-swe-agent achieves 65% on SWE-bench verified in 100 lines of Python, matching SWE-agent performance while being significantly simpler and more hackable. Development effort has shifted to mini-swe-agent. SWE-agent remains available and maintained under the SWE-agent organization but new features and improvements are landing in mini-swe-agent first. Integration depth is minimal by design. SWE-agent connects to GitHub repositories for issue reading and PR submission. It uses whatever LLM the user configures as the reasoning engine: GPT-4o, Claude Sonnet 4, or any compatible model. The entire agent is governed by a single YAML configuration file. There are no IDE extensions, no project management integrations, no Slack or collaboration tool connections, and no MCP support. This is deliberate: SWE-agent is research infrastructure designed for maximal LLM agency with minimal scaffolding, not a productized developer tool. Teams needing IDE integration should evaluate Cursor ($16/mo billed annually) or Cline (free, open source). Teams needing enterprise collaboration integrations should evaluate OpenAI Codex or Devin. Pricing follows the open-source BYOK model: SWE-agent is free under MIT license, with users paying only for LLM API usage through their chosen provider. Typical runs cost $1 to $10 per issue depending on complexity and chosen model, materially cheaper than commercial autonomous coding tools at moderate use volumes. There are no subscriptions, no credits, and no hosted service. SWE-agent is not the right fit for teams expecting a productized commercial tool. It is research infrastructure with no commercial support, SLA, compliance certifications, or polished UX. Setup requires Python environment configuration, API key management, and command-line expertise. Organizations needing enterprise-grade autonomous coding should evaluate Devin ($20/mo) or Claude Code ($17/mo billed annually) instead. SWE-agent is best suited for academic researchers, security auditors, and developers comfortable with research-grade tooling who want to understand exactly how autonomous coding agents work. Current state Q3 2026: SWE-agent has been superseded by mini-swe-agent for general use, though the original repository remains maintained. SWE-agent-LM-32b achieved state-of-the-art results on SWE-bench in May 2025. The project continues to be used for offensive cybersecurity research, competitive coding challenges, and custom autonomous tasks beyond standard issue-fixing. No security certifications exist. The project is MIT licensed and fully open source.