Open-source Devin alternative achieving 12.3% on FULL SWE-bench (vs Devin's 13.84% on 25% of test set). Works through specialized terminal allowing file opening/scrolling/searching, line editing with syntax check, and test writing/execution. Custom-built interface critical for performance. Uses Agent-Computer Interface (ACI) constraining commands to run/look/edit code and submit to GitHub. Code undergoes linter checks with feedback loop. Reads 100 lines at a time for better LM comprehension. Released by Princeton NLP team.