A leech is a card that has failed enough times that the scheduling has given up on it, and the fix is almost always rewriting the card rather than reviewing it again — which is exactly the sort of judgement a model can help with. Writing a dated tag back is what keeps that useful over weeks: you can tell at a glance which leeches you have already reworked. Anki has to be open for any of it to work.
A narrow Anki server aimed at one problem: the cards that keep failing. It talks to a running Anki through the AnkiConnect add-on, retrieves cards tagged as leeches with enough data to reason about why they are hard, and writes a date-stamped tag back so you can see which ones you have already been through.
- `get_leech_cards` retrieves cards tagged as leeches, with a `detailed` flag for full card data or just IDs, and a `count` for a random sample
- `tag_reviewed_cards` writes a date-stamped tag onto card IDs you pass, in the form `見直し::YYYYMMDD`
- `custom_tag_prefix` replaces that prefix if you keep your collection in another language
- A mock mode for trying it without touching a real collection
Anki running with the AnkiConnect add-on installed — the server talks to it over HTTP, not to the collection file. Node.js and npm to build: `npm install`, `npm run build`, then point your client at `dist/index.js` (package `anki-mcp-server`, version 1.0.1). `ANKI_CONNECT_URL` defaults to `http://localhost:8765` and `ANKI_CONNECT_VERSION` to `6`; if localhost does not connect, set the URL to your machine's LAN address and allow it in AnkiConnect. `ANKI_MOCK_MODE` runs it without Anki at all.
Build from source — clone the repository and build it, then point your client at the binary