SilvaHerald/anki-mcp-server

📚 NestJS Anki MCP Server

An MCP server built with NestJS that connects to AnkiConnect and enables AI-assisted flashcard generation and management.

This project exposes:

• MCP Prompts → generate effective cards (Basic & Cloze) using AI, following active recall & spaced repetition principles.
• MCP Tools → create cards directly in Anki via AnkiConnect.

🎯 Goal: make vocabulary learning easier by combining local LLMs (via Ollama/DeepSeek/Qwen/etc.), MCP prompts, and Anki.

✨ Features

• 🧠 Prompt-based card generation

  • Define vocabulary words/phrases
  • Get structured Anki-ready JSON cards (Basic + Cloze)
  • Supports tagging & deck mapping (Words, Idioms, Collocations, Phrasal Verbs, Slangs)

• 🛠 MCP Tools

  • Create Anki cards programmatically
  • Supports Basic and Cloze models
  • Bulk note insertion via AnkiConnect

• 🔗 Seamless AI integration

  • Works with local LLMs via Ollama (deepseek-r1, qwen2.5, llama3.1, …)
  • Or swap in a cloud model (OpenAI, Anthropic, Gemini)

• 🧩 NestJS + @rekog/mcp-nest

  • Strong typing with Zod schemas
  • Clean modular design (PromptsService, ToolsService, AnkiConnectService)

🚀 Getting Started

1. Prerequisites

• Anki desktop app
• AnkiConnect plugin installed in Anki
• Node.js ≥ 22
• pnpm ≥ 9
• (Optional) Ollama for local LLMs

2. Clone & install

git clone https://github.com/<your_username>/anki-mcp-server.git

cd anki-mcp-server

pnpm install

3. Run MCP server

pnpm run start:dev

This will start a NestJS MCP server exposing:

• Prompts (e.g., get-effective-cards-prompt)
• Tools (e.g., create-anki-cards)

4. Connect to a client

• Use the MCP Inspector to test

pnpm run mcp:inspect

• Or connect via ChatGPT / Claude Desktop if you expose it as a Streamable HTTP MCP server.

5. Example usage

Generate cards via MCP Prompt

{
  "cards": [
    {
      "deck": "Languages::English::New Words",
      "model": "Basic",
      "fields": {
        "Front": "They burned the letter to hide the evidence.\nWhat does the word 'burn' mean here?",
        "Back": "Definition (EN): to set something on fire\nDefinition (vi): đốt\nExample: The chef burned the toast.\nSynonyms: ignite, scorch"
      },
      "tags": ["english", "new_word"]
    },
    {
      "deck": "Languages::English::New Words",
      "model": "Cloze",
      "fields": {
        "Text": "They {{c1::bur}}ned the letter to hide the evidence.",
        "BackExtra": "They burned the letter to hide the evidence."
      },
      "tags": ["english", "new_word"]
    }
  ]
}

Insert into Anki via MCP Tool

{
  "deck": "Languages::English::New Words",
  "front": "They burned the letter to hide the evidence.\n\nWhat does the word 'burn' mean here?",
  "back": "Definition (EN): ...",
  "tags": ["english", "new_word"]
}

🛠 Project Structure

Read more at .

🔄 Architecture

Read more at

🧪 Development Notes

• Always validate AI output with Zod schemas before sending to AnkiConnect.

• Use temperature: 0 and structured JSON mode in Ollama/OpenAI for consistent results.

• Supported Anki models:

  • Basic: fields → Front, Back

  • Cloze: fields → Text, Back Extra

🤝 Contributing

PRs, issues, and discussions are welcome! Feel free to fork, adapt prompts/tools, or add support for more Anki models.

📄 License

MIT — free to use, modify, and distribute.

🙌 Acknowledgements

• Model Context Protocol (MCP)
• NestJS
• MCP-Nest
• Anki + AnkiConnect
• Ollama and open-source LLMs (DeepSeek, Qwen, LLaMA, ...etc)