SWE-agent

What it is for

SWE-agent is Princeton’s open-source software-engineering agent: you give it an issue, it edits a repo, and the Inspector shows each trajectory. Use it when the job is a research-grade agent you run yourself.

The software is MIT and $0. You bring a model key. The cover is the official Trajectory File Viewer — run list, pass/fail, step counts.

What makes it unique: the Inspector is a real product UI for SWE-bench-style runs, next to the CLI you start from docs.

Key features

• Inspector — trajectory list with submitted / pass-fail status and step counts (cover).
• CLI agent — sweagent run against a GitHub issue, a local repo, or a markdown problem file (docs).
• SWE-bench loop — evaluate patches the same way the paper does.
• Docker / Modal env — run the agent in a documented image or a cloud deployment.
• Patch hand-off — apply a successful patch locally, or open a PR from a GitHub issue run.
• YAML config — split agent, env, and problem settings across --config files.

Who it's for

Best fit: researchers and platform engineers who want to run and inspect coding-agent trajectories on their own machines.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from swe-agent.com and github.com/SWE-agent/SWE-agent (MIT).

• CLI / Inspector — $0. MIT. You install and run it.
• Tokens — billed by the model vendor you connect.
• Hosted SWE-agent seat — the pages we fetched list the open-source project.

Limitations / costs

• $0 is the software. A long SWE-bench run still spends vendor tokens.
• You operate the stack. Docker, keys, and trajectories stay on your side.
• Cover is the Inspector list, the public UI for those runs.
• Re-check the repo if a hosted SKU appears later.

Alternatives

• OpenHands — OSS + hosted automations; Free SaaS Individual.
• Aider — git-native TUI; Apache-2.0 $0.
• OpenCode — MIT TUI/desktop/IDE; software $0.

Official tutorial

Written SWE-agent docs plus the CLI tutorial.

Website

https://swe-agent.com

Repository

https://github.com/SWE-agent/SWE-agent

Source

Catalog row curated for AI Tools Nav. Cover is the official SWE-agent Inspector trajectory list (pass/fail + step counts). Software $0 MIT; tokens billed by the vendor you connect.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.