gpt-oss:20b passes app-building tests in a Claude Code plugin
agent-smith is a helper for Claude Code that sends heavy draft work to free models or , so Claude can spend more of its usage on review and judgment. In a personal test setup, OpenAI’s passed 14 tasks twice in a row, covering code writing, information extraction, repository edits, and full app builds. Under that test rule, it became the first model trusted for app-building work.
The 13 GB model ran on a MacBook Pro M3 with 36 GB of memory and built , a CSV tool, and an HTTP API that passed . It also succeeded on the same tasks where other models kept failing. It did worse than gemma4:26b in a blind design comparison.
The main code-quality issues were leftover debug output, unused code paths, and text that described errors the code never actually raised.
Key points
- agent-smith splits Claude Code work across free models and .
- passed 14 personal evaluation tasks twice in a row.
- It built , a CSV tool, and an HTTP API that passed .
- gemma4:26b beat it on design quality in a blind comparison.
- The model still left code-cleanliness problems such as unused branches and misleading .