A native multi-language display layer for Claude Code.
The model keeps reasoning and replying in English (best reasoning, no CJK
"token tax"); cc-lingua translates only the on-screen narration into your
language. Code, file paths, shell commands, URLs and markdown are preserved
verbatim. It runs as a MessageDisplay hook that fails open — if anything
goes wrong, you just see the original English.
This is the open-source, bring-your-own-key edition: no server, no account, no subscription. No API key is bundled — you buy your own from DeepSeek (a paid third-party service, cents per day) and the hook calls DeepSeek directly with it, stored locally. ~120 lines of stdlib Python, nothing to run in the background.
中文用户：纯开源版，无服务器 / 账号 / 订阅。本项目不含 key —— 你需要自己到 DeepSeek 官网购买一个 API key（第三方付费服务， 每天通常只花几分钱），翻译直接用你自己的 key。安装见下方「快速安装」。
You need Claude Code and a DeepSeek API key (get one at https://platform.deepseek.com/api_keys and top up a few dollars — translation is very cheap).
Then just paste this to Claude Code and let it install itself:
Clone https://github.com/partyfly/cc-lingua and follow its INSTALL.md to install cc-lingua for me.
Claude Code will clone the repo, read INSTALL.md, write the hook,
pick your display language (the one you're chatting in — or it asks), take your
DeepSeek key, register the hook, and turn it on. Because settings.json hooks
hot-reload, its very next reply appears in your language.
把上面这句（替换成你的仓库地址）贴给 Claude Code 即可；它会读
INSTALL.md自动装好， 中途会向你要 DeepSeek key。装完当前这条回复就变中文。
Claude Code ──(English narration)──► MessageDisplay hook ──► DeepSeek ──► your language
│ translate.py on screen only
└── transcript & model context stay English (full reasoning, no token tax)
- Display-only. The saved transcript and everything the model sees stay English. Only the rendered text on your screen is swapped.
- Fail-open. Toggle off, non-final chunk, empty text, already-in-your-language, missing key, network error → the hook prints nothing and you see the original English. It never blanks or blocks a message.
- Code-safe. The translation prompt reproduces code blocks, inline code, paths, commands, URLs, numbers and identifiers verbatim and preserves markdown.
- Cached. Identical strings are translated once and reused from
~/.claude/cc-lingua/cache/, so repeats cost nothing.
~/.claude/cc-lingua/config.json (see client/config.example.json):
{
"api_key": "sk-your-deepseek-key",
"base_url": "https://api.deepseek.com",
"model": "deepseek-v4-flash",
"target": "Simplified Chinese"
}| field | meaning |
|---|---|
api_key |
Your DeepSeek key. Or set env DEEPSEEK_API_KEY (takes precedence). |
base_url |
OpenAI-compatible endpoint. DeepSeek by default; any compatible provider works. |
model |
deepseek-v4-flash — cheapest tier, the right default. deepseek-v4-pro for higher quality (~3x price). Don't use deepseek-chat; it's deprecated 2026-07-24. |
target |
Any language as free text: "Simplified Chinese", "Japanese", "Korean", "Spanish", "French", … |
rm ~/.claude/cc-lingua/enabled # pause (display back to English)
touch ~/.claude/cc-lingua/enabled # resumeTo uninstall, paste to Claude Code: "follow UNINSTALL.md" — or manually
rm -rf ~/.claude/cc-lingua and remove the hook from ~/.claude/settings.json.
See UNINSTALL.md.
- Cost.
deepseek-v4-flashis ~$0.14 / 1M input tokens, ~$0.28 / 1M output. Typical Claude Code sessions translate to cents per day; the local cache cuts repeats to zero. You pay DeepSeek directly. - Privacy. No telemetry, no middleman. The only network call is from your
machine to your configured
base_url, sending only the narration text Claude Code was about to show you. Yourconfig.json(with the key) is git-ignored and never leaves your machine.
~/.claude/cc-lingua/hook.log has one line per event. FAIL:api almost always
means a bad key, no account balance, or a wrong model / base_url. skip:nokey
means config.json has no api_key.
MIT — see LICENSE. Contributions welcome.