1. You waste tokens — non-English text takes 2-3x more tokens than English for the same meaning. Every prompt, every response, every turn in context is inflated.

2. Claude reasons worse — it spends context budget on language switching instead of actually thinking about your code.

I built claude-ts to fix this. It's a translation proxy that sits in front of Claude Code:

You (any language) → Haiku/Ollama (→ EN) → Claude Code (EN) → Haiku/Ollama (→ your lang) → You

Claude Code always works in English internally — better reasoning, fewer tokens. The translation costs almost nothing (Haiku) or literally nothing (local Ollama).

pip install claude-ts

- 8 languages supported (ko, ja, zh, th, hi, ar, bn, ru) - Free local translation via Ollama - Real-time agent tree visualization - All Claude Code features preserved