anki-agent
# anki-agent
I want to turn my study notes into Anki cards with Codex or Claude Code, without having an agent click through Anki every time.
This project gives AI agents a simple way to work with Anki through MCP (Model Context Protocol). Anki stays open in the background, while a local add-on and MCP server let the agent work with the collection through Anki's own API.
The add-on checks the connection, lists and creates decks, lists note types, searches notes, and previews and creates Basic and Cloze notes in batches. It also previews and deletes cards or decks, but only after you click Yes in Anki. A recovery backup must finish before anything is deleted. The connection is local and authenticated, with fourteen defined tools. Decks can be renamed or moved in the hierarchy, and selected cards can move between normal decks while keeping their IDs and scheduling.
The MVP has been tested with Anki 25.9.2 in a disposable profile, including a real MCP client. Preview shows proposed note content; it does not render card templates. Media and other note structures are outside the first version.
## Default study prompt
The [`study_prompts` folder](study_prompts/) includes a [default prompt for turning course material into Anki Cloze cards](study_prompts/course-material-to-anki-cloze.md). Fill in the deck, topic, assessment scope, and sources before using it with your agent.
[Setup for Codex and Claude Code](docs/setup.md) · [Deck organization](docs/deck-organization.md) · [Deletion and backups](docs/deletion-and-backups.md) · [Local text example](examples/from_text.py) · [Architecture](docs/architecture.md) · [Testing and packaging](docs/development.md)
Licensed under [AGPL-3.0-or-later](LICENSE). No public release or AnkiWeb upload yet.
An independent project, not affiliated with Anki, OpenAI, or Anthropic.
## Editing existing answer fields (0.4.0)
`update_note_fields` edits 1-100 existing notes in place after user authorization. Inspect the current note IDs and HTML fields with `search_notes`, then supply plain-text replacements and matching `expected_fields` containing the exact current HTML. Only non-first fields can be edited, protecting card-generating text. Changes use a single undoable Anki transaction; note/card IDs, tags, decks, and scheduling are preserved. Stale fields abort the entire batch. Identical retries are safe. Restart Anki after installing an upgraded add-on, with the user’s approval.
TDQS
Scored across 15 tools
Each tool targets a distinct resource+action, and the preview/apply pairs (preview_notes/create_notes, preview_delete_cards/delete_cards, preview_delete_deck/delete_deck) are explicitly framed as dry-run versus commit steps, so an agent can tell them apart without guesswork. Card-level versus deck-level deletion is also clearly separated by scope.
Nearly every tool follows a snake_case verb_noun pattern (list_decks, create_deck, rename_deck, move_cards, search_notes, delete_deck), including the compound preview_* prefix. The lone outlier is connection_status, a noun-only status check, which is a minor but visible deviation.
At 15 tools the surface is dense but every tool maps to a real operation (connection check, deck CRUD, note lifecycle, card moves/deletes, backup), with no redundant wrappers. Nothing feels padded or missing from the count perspective.
Core lifecycle is well covered: decks (list/create/rename/delete), notes (search/preview/create/update), cards (move/preview-delete/delete), plus backup. Gaps remain around tags (no tag add/remove/list), card scheduling controls (suspend/bury/reset), and direct single-note deletion, which agents must work around via card deletion.