I translate work emails every day in a language I'm still learning. For two years, I'd paste a sentence into Google Translate, get one answer, send it, and never know if I'd used a register that made me sound like a child or overly stiff. I wasn't learning anything — I was just getting through the task.
So I built konid: an MCP server that returns three translations per query, ordered casual to formal, with the register explained and cultural context on why the options differ.
The audio piece matters more than I expected. When you're learning a tonal language or one with vowel-length distinctions — Mandarin, Japanese, Arabic — reading the romanization and hearing the actual tone contour are completely different cognitive inputs. I kept mispronouncing Japanese words I'd read dozens of times until I heard them. konid plays pronunciation directly through your speakers via node-edge-tts, no external API key required.
The three-option structure forces comparison you wouldn't otherwise do. For "I missed you today" in Japanese, the casual form uses a completely different verb construction than the polite form — not just a suffix swap. Seeing them side by side with an explanation of when each is appropriate is the kind of thing a tutor would tell you, not a dictionary.
Install into Claude Code:
claude mcp add konid-ai -- npx -y konid-ai
Works with Cursor, VS Code Copilot, Windsurf, Zed, JetBrains, and Claude Cowork. Also runs as a ChatGPT app via Developer mode with the endpoint https://konid.fly.dev/mcp.
13+ languages: Mandarin, Japanese, Korean, Spanish, French, German, Portuguese, Italian, Russian, Arabic, Hindi, and more.
MIT licensed. https://github.com/robertnowell/konid-language-learning
Top comments (0)