Translate Proxy

translate-proxy

A transparent, universal translation relay for AI coding agents. You type in your language — the model thinks in its language — the answer comes back in your language.

Sits between an AI client (Claude Code, opencode, Hermes, …) and the model endpoint it already uses:

• ingress — your USER_LANG prompt (and history prose) is translated to MODEL_LANG before the model sees it (better reasoning, fewer tokens);
• egress — the model's MODEL_LANG answer is buffered, translated back to USER_LANG with a cheap translator model, and streamed to you in the original wire format.

Python standard library only. Works for **any language pair** (`USER_LANG`/`MODEL_LANG`).

✨ You need this

*You* write in *your* language. The model *always* thinks in English. The answer comes back in *your* language. The English in between is invisible — and nearly free.

• 🗣️ Ask in Polish, German, Japanese, Arabic, Hindi, Swahili… ~100 languages, zero setup.
• 🧠 The model only ever sees and produces English — its strongest language. Better reasoning, and you never waste tokens on "sorry, could you say that in English?".
• 💸 Save tokens — say it once, in your language, done.
• 📉 Save 40%. 📈 Save 140%. (Math says you can't save more than 100%, but our heart says otherwise.)
• 🤖 Works with Claude Code, opencode, Hermes — anything that speaks the Anthropic or OpenAI wire formats.

How the magic happens

you (PL):       "Czesc, jak sie masz?"
   └─ proxy ──► model (EN): "Hi, how are you?"
   └─ model ──► proxy:      "I'm doing well, thank you!"
   └─ proxy ──► you (PL):   "U mnie wszystko dobrze, dziękuję!"

Your prompt is translated to English **once** (going in), and the model's answer is translated back **once** (coming out). That's the whole cost.

💰 What you actually pay

The main model costs what your provider charges. **The only *extra* cost is the cheap translation model** — one OpenRouter-style call per