I'd been translating work emails in Japanese for months — reading the same polite closing phrases daily, copy-pasting from Google Translate, and still blanking on them when I needed to write from memory. The problem wasn't vocabulary. It was that I'd only ever read the words. Japanese pitch accent and vowel length mark the difference between words that look identical on the page; without hearing them, you're building a half-formed mental model that doesn't stick.
So I built konid partly to fix my own Japanese, and the audio piece turned out to matter more than I expected.
When you ask konid how to say something, it returns three options ordered casual to formal — not just the translation, but an explanation of why one register fits a colleague and another fits a friend. Then it plays pronunciation through your speakers via node-edge-tts, no external API key. Hearing the pitch drop on a Japanese keigo phrase or the vowel length in Arabic that changes word class is what moves something from "I've seen this" to "I actually know this."
The 13-language support includes Mandarin, Japanese, Korean, Arabic, and Hindi — languages where tone or length distinctions are phonemic, meaning a flat text translation is actively misleading about how the word sounds in real use.
Installs as an MCP server in one line:
claude mcp add konid-ai -- npx -y konid-ai
Works in Cursor, VS Code Copilot, Windsurf, Zed, JetBrains, and Claude Cowork. Also available as a ChatGPT app via Developer mode with endpoint https://konid.fly.dev/mcp.
MIT licensed. github.com/robertnowell/konid-language-learning
Top comments (0)