Language learning MCP servers: what each one adds to your AI

A language learning MCP server gives your AI something a chat can't do on its own. That might be a voice, a microphone, a set of exercises, or a word list that's still there next week. The AI still does the teaching. The server gives it hands.

Search for one and most of what comes back is about programming languages, or about MCP in general. So here's the actual list as of October 2026, what each one does, and how to pick. Full disclosure up front: one of them is ours, and I'll say which.

What an MCP server adds to an AI tutor

Claude and ChatGPT are already good at explaining a sentence or fixing your writing (what Claude does well, and what it forgets, and the ChatGPT modes worth using). They fall short on anything that has to live outside the chat:

• Sound. Voice mode can chat with you, but it won't make audio flashcards you keep, or check your pronunciation word by word.
• Memory. A chat doesn't keep the words you looked up, or know which ones are due today.
• Structure. Quizzes and drills work better as real widgets than as walls of text.

Every server below covers one of those three. None of them replaces the model. They plug into it.

The servers, as of October 2026

Mossglen (ours). A word list and review schedule your AI can read and write. You save words while reading web pages in the Mossglen extension, or straight from the chat. Saying "help me review today's words" gets you a quiz on the ones that are due. The AI also gets your level, so explanations are pitched to it, and you can take a level test in the chat. It's remote, so nothing runs on your computer. Works in Claude (web, desktop, phone), ChatGPT (web), Claude Code, Codex, Cursor, VS Code and Gemini CLI. Needs a free Mossglen account. Eight languages.

language-learning-mcp. An MCP App for claude.ai that draws exercises right in the chat: flashcards, multiple choice, dictation with audio, sentence builders, matching, fill in the blank, and a DELF or TOEFL style test. You add one URL in Claude's settings, and Claude grades your answers. It's practice in the moment. The README doesn't mention keeping a word list.

langlearn-tts. Gives Claude Desktop a voice. Ask it to pronounce a word, make audio flashcards, or merge a list of words into one audio file, in 70+ languages. You bring your own key from ElevenLabs, OpenAI or AWS Polly, so the audio costs whatever that provider charges. It also ships 28 tutor prompts, one for each mix of seven languages and four levels.

mcp-server-pronunciation. A local voice coach for English. You speak, it transcribes on your machine with Whisper, and the AI gets pronunciation, grammar and fluency feedback in the same turn. There's a drill mode for working on one sentence. Audio never leaves your computer. The learner hints are tuned for Korean speakers so far. Needs Python 3.11 and a microphone.

konid. An expression coach. Tell it what you want to say and it gives you three ways to say it, casual to formal, with notes on tone and audio. Runs as a Claude Cowork plugin on a Pro or Max plan, or in Claude Code, where it needs Node.js and an Anthropic API key.

Polyprompt. For programmers. You write coding prompts in your own language, it translates them into the one you're learning before the agent runs them, and the new words go into a spaced repetition deck. Works with Claude, Codex and Gemini on the subscription you already have.

One that isn't an MCP server but turns up in the same searches: Fluent, a Claude Code plugin that makes the terminal a tutor with its own progress files. Good if you live in Claude Code. It doesn't connect to anything outside it.

Which one to connect

You want	Connect
Your saved words reviewed in the chat	Mossglen
Interactive exercises inside claude.ai	language-learning-mcp
Claude to say words out loud	langlearn-tts
Feedback on how your English sounds	mcp-server-pronunciation
Natural ways to say one specific thing	konid
Language practice while you code	Polyprompt

They don't conflict, so connecting two is fine. A sensible pair is one for sound and one for memory.

How to connect one

There are two kinds, and setup differs:

• Remote servers (Mossglen, language-learning-mcp) are just a URL. In Claude, go to Settings, then Connectors, then Add custom connector. In ChatGPT, it's Settings, then Apps, with developer mode on. ChatGPT's custom connectors work only on the web for now, and depend on your plan. In Claude Code it's one claude mcp add command.
• Local servers (langlearn-tts, mcp-server-pronunciation, konid) run on your computer. They work in desktop apps and coding tools, not on claude.ai in a browser or on your phone. Follow each README.

Our steps for every app are on the Mossglen AI page.

Before you connect anything

• Check what it can see. A connector reads and writes data for you. Ours sees your learning language, level, word list, reading library and a short summary of your learner profile. Not your password, not your payment details. Ask the same question of any server you add.
• Check who pays. Servers that call a speech or AI service with your own API key bill you per use. Ones that only store data, or borrow the chat's own model, run on your normal Claude or ChatGPT allowance.
• Check it's alive. Most of these are one-person projects with a handful of GitHub stars. Look at the last commit before you build a habit on one.

Where this fits

An MCP server doesn't make the AI a better teacher. It makes it a teacher that remembers, speaks, or hands you a worksheet. Which jobs to give that teacher in the first place is its own question (how to use AI to learn a language). The input still has to come from somewhere, and for most learners the best source is real text at the right level (how to learn a language by reading). The server is how what you read follows you into the chat.