I was writing a work email in Japanese — still learning, good enough to get through it, not good enough to know whether I was being rude. Google Translate gave me one sentence. No register note. No indication whether I'd just addressed a senior colleague like a peer.
konid returns three options per query, ordered casual to formal, with the register explained and cultural context on why the choices diverge. For Japanese, that difference is structural — keigo isn't just polite, it's a whole conjugation system — and a literal tool won't tell you that.
The MCP setup is a single command:
claude mcp add konid-ai -- npx -y konid-ai
That puts it in Claude Code. The same server also works in Cursor, VS Code Copilot, Windsurf, Zed, JetBrains, and Claude Cowork. If you're on ChatGPT instead, there's a Developer mode path using the endpoint https://konid.fly.dev/mcp — same responses, same three-option format, different client. Write the MCP logic once; all four client surfaces pick it up.
Audio pronunciation runs through node-edge-tts, which means no external API key and no separate voice service to configure. It plays through your speakers directly.
The language list covers 13+ including Mandarin, Japanese, Korean, Spanish, French, German, Portuguese, Italian, Russian, Arabic, and Hindi.
The part that made me actually build this rather than keep pasting into a browser tab: every query teaches you something. The nuance comparison between options is where the language learning happens — not just "here's the translation," but "here's why you'd pick this one over that one."
MIT licensed. https://github.com/robertnowell/konid-language-learning
Top comments (0)