SWE-agent

SWE-agent (Princeton NLP, first released April 2024, presented at NeurIPS 2024) takes a GitHub issue and an LLM of your choice, and lets the model browse, edit, and execute code inside a sandboxed Docker container through a purpose-built command set the authors call an Agent-Computer Interface (ACI) — structured file-viewing, editing, and search commands designed to be easier for an LLM to use reliably than a raw shell. That ACI framing was the project's genuine research contribution and is documented in the NeurIPS paper. (https://github.com/SWE-agent/SWE-agent, https://collaborate.princeton.edu/en/publications/swe-agent-agent-computer-interfaces-enable-automated-software-eng/)

The docs site (swe-agent.com) currently shows a banner reading "SWE-agent is now in maintenance-only mode" and a popup recommending mini-swe-agent instead, with the stated reason "Same performance, much more simple." This isn't a rumor or a stale note — it's the live, vendor-authored front page of the docs as of today. GitHub activity backs up the "maintenance" label: SWE-agent/SWE-agent still received bug-fix and doc-correction commits through July 2026 (most recently merged 2026-07-16), and the repo was pushed to as recently as 2026-08-17, but the last tagged release is v1.1.0 from May 2025 — over a year with no new version, while commits keep landing on main. Read that as: bugs get fixed, nothing new ships. mini-swe-agent, from the same GitHub org, is under active version releases (v2.0.0 as of this year) and is what the team is putting effort into now. (https://swe-agent.com/latest/, https://github.com/SWE-agent/SWE-agent/commits/main, https://github.com/SWE-agent/SWE-agent/releases)

On the official SWE-bench leaderboard (swebench.com), the harness labeled "mini-SWE-agent" appears in 107 separate result rows — used by Anthropic, OpenAI, Google DeepMind, Moonshot, Z-AI and others to report official SWE-bench Verified scores for their models — with resolve rates up to 76.8% for the strongest model/harness pairing found today. The original "SWE-agent" harness has 20 rows on the same leaderboard, reflecting its earlier, lower-ceiling era (the 2024 launch figure widely cited was 12.29% resolved with GPT-4). In short: the ACI concept SWE-agent introduced won, but the numbers you'd want to quote today belong to the successor project, not this one. (https://www.swebench.com/)

Free and MIT-licensed (confirmed via the GitHub API license field), runs entirely on infrastructure you control, and has no vendor pricing page because there's no vendor — you pay only for the LLM API calls it makes on your behalf and, if you use the default setup, for Docker container time. There is no hosted or managed offering; this is a self-run CLI/library. (https://api.github.com/repos/SWE-agent/SWE-agent)

The sweagent package on PyPI is real (published by the actual author, John Yang) but stuck at version 0.0.1, uploaded 2024-04-02 and never updated since — while the GitHub project has moved through v1.1.0 and well beyond in active commits. pip install sweagent will not get you a working, current copy of the tool; the project's own install docs point to cloning the repo and running pip install --editable . from source instead. (https://pypi.org/pypi/sweagent/json, https://swe-agent.com/latest/installation/source/)

A related sub-project, EnIGMA, extends the same ACI approach to offensive security / CTF challenges, adding "Interactive Agent Tools" for things like concurrent debugger and shell access, and the vendor reports it solved 13.5% of NYU CTF benchmark challenges, said to be over 3x the prior best agent on that benchmark. This is a vendor-reported figure from the docs, not independently corroborated here, and is a narrower research use case than the core issue-resolution tool. (https://swe-agent.com/latest/background/)