anki automation agents

📝 Anki Automation Agents

🚀 My simple CrewAI-based automated workflow to collect Japanese vocabulary, craft high-quality bilingual notes, and add them directly into Anki via an external MCP server.

⚙️ CrewAI Agents

👨‍💻 Agents

• definition_writer: JP→VI Lexicographic Note Crafter

Produces clean Vietnamese meanings, usage notes, natural JP examples, and contrasts with easily confused terms.

• anki_operator: Anki MCP Operator

Orchestrates the note creation flow, checks duplicates, creates decks if missing, and pushes notes into Anki through MCP.

🔧 Tasks

• compose_back_content_task:

Input: a headword (JP)

Output: learner-friendly “Back” content (VN gloss, usage, 📌 examples, and near-miss terms).

• create_anki_note_task:

Full end-to-end automation:

1. Ensure deck exists (or create)

2. Check duplicates

3. Add note with {Front: word, Back: content} via MCP

📚 MCP Server

MCP Server: simple-anki-mcp-server

🚀 Run agent

Enviroment Setup (uv)

uv sync

Run crew

After setting up the enviroment we can run the agent using command:

crewai run

Remarks

CrewAI uses its own implementation of ReAct so i saw some injected prompt related to React, take a look at CrewAI React Prompt. Therefore, some LLM models like gpt-oss-120b might not be optimized with this prompts and is needed to be tuned.

CrewAI may not support the Function Calling Strategy (the function calling or tool calling which provides ~100% accuracy in params and function name features that vllm and openai offer), so we need to custom new agent executors on top of the frameworks.