A practical test of CLI AI coding agents in real codebases
More than 15 CLI were tried for months on real production repositories. The test was not a simple demo app, but projects with about 200 files, older coding patterns, and tests that had to pass. The work covered Go , TypeScript monorepos, and Python .
The main question was not which tool wrote the nicest code, but which one could be given a task and later leave behind a working commit. Most tools failed when the task needed judgment, invented file paths that were not real, or broke tests without catching the problem. Only 3 tools passed that practical bar, and Claude Code was ranked first.
Claude Code was described as the safest choice for refactoring work that touches many files. The listed reasons include 96,000 GitHub stars, more than 250 releases, Opus 4.8, and an 88.6% score on , which is presented as a test of solving real end to end.
Key points
- The comparison covered more than 15 CLI .
- The test used real production repositories, not small demo projects.
- Most tools failed on unclear choices, fake file paths, or broken tests.
- Claude Code was ranked first for work.
- Claude Code was tied to Opus 4.8 and an 88.6% score.