ankix

Turn words you look up while reading, watching, or browsing into contextual Anki translations.

Reading on a Kindle and looking up words builds a vocabulary list nobody reviews. Watching foreign-language YouTube with subtitles surfaces words you want to remember and then loses them the moment the video ends. Reading an article in another language means alt-tabbing to a dictionary for every other word. ankix closes those loops: it reads your Kindle's vocab.db, a video's subtitle track, or a web article, defines each word with a local Ollama model using the sentence it appears in — not just a bare dictionary entry — and syncs ready-to-study notes straight into Anki via AnkiConnect.

Install

brew install --cask joshgummersall/ankix/ankix

Requirements

• Ollama — the kindle, youtube, and web commands all use the same ankix model (ollama/vocab/Modelfile)
• Anki running with the AnkiConnect add-on installed (Tools → Add-ons → Get Add-ons..., code 2055492159, then restart Anki)
• yt-dlp for the youtube command

After installing, build the local Ollama model once:

ankix install

ankix kindle — Kindle vocabulary builder

ankix kindle vocab ~/Documents/vocab.db --lang en --deck "Kindle Vocab"

Find vocab.db on your Kindle at system/vocabulary/vocab.db, or in a backup/export of the device.

Each note uses Anki's built-in Basic note type: Front is the highlighted sentence from the book with the looked-up word in bold; Back is the definition.

Positional argument: path to vocab.db.

Flags:

• --lang — language prefix to filter words by, e.g. en, es (default en)
• --deck — Anki deck to sync into (default Kindle Vocab)
• --model — Ollama model used to define words (default ankix)
• --tag — tags applied to new notes (default AnkiX::Source::Kindle)
• --dry-run — preview without writing to Anki
• --ankiconnect-url — AnkiConnect endpoint (default http://localhost:8765)

Only words not already marked Mastered in vocab.db are considered, and any word that ends up in Anki (added, or already there) is marked Mastered (sets WORDS.category to 1), removing it from the Kindle's Vocabulary Builder review queue — that's what tracks sync progress across runs, no separate watermark is kept. This opens vocab.db read-write (except for a headless --dry-run), so point it at the device itself rather than a copy if you want the change to take effect on the device. Before writing anything, vocab.db is snapshotted into a timestamped backup log under $XDG_CONFIG_HOME/ankix/kindle-backups (or the OS equivalent), namespaced per source file. Use ankix kindle vocab db list <vocab.db> to see past snapshots and ankix kindle vocab db restore <vocab.db> <index> to roll back to one (which itself snapshots the current file first, so a restore is always undoable).

The namespace is keyed by the vocab.db path you pass in, not the device itself — vocab.db has no serial number or account ID to key off. If the same Kindle ever mounts at a different path, or you copy vocab.db somewhere new, list/restore will see it as an unrelated file and start a fresh backup history rather than continuing the old one. Always point list/restore at the same path you've been syncing that device with.

Re-running sync checks AnkiConnect for an existing note with a matching headword in the target deck to skip words already synced, so it's safe to re-run.

Definitions are hydrated through the dict.Provider interface (internal/dict/dict.go), so other sources can be added later without touching sync logic. The only implementation today is internal/dict/ollama.

ankix youtube — YouTube transcripts

ankix youtube fetch <youtube-url>
ankix youtube review <transcript-file.vtt>

fetch downloads subtitles via yt-dlp and opens the transcript in a terminal UI for browsing and generating cards; review opens an existing .vtt transcript file directly, skipping yt-dlp.

Flags (persistent across both subcommands):

• --deck — Anki deck name (default AnkiX)
• --ankiconnect-url — AnkiConnect URL (default http://localhost:8765)
• --ollama-url — Ollama URL (default http://localhost:11434)
• --ollama-model — Ollama gloss model name (default ankix)
• --sub-lang — subtitle language code (default es)
• --cache-dir — subtitle cache directory
• --no-gloss — skip Ollama gloss lookups

ankix web — Web articles

ankix web fetch <url>

Fetches and extracts the article text from a URL, then opens it in a terminal UI for browsing and generating cards — the same TUI used by ankix youtube.

Flags:

• --deck — Anki deck name (default AnkiX)
• --ankiconnect-url — AnkiConnect URL (default http://localhost:8765)
• --ollama-url — Ollama URL (default http://localhost:11434)
• --ollama-model — Ollama gloss model name (default ankix)
• --no-gloss — skip Ollama gloss lookups

Using a different language

Everything except the Ollama model itself is language-agnostic — the --lang/--sub-lang flags just select a language code, and Kindle/subtitle parsing don't assume any particular language. ollama/vocab/Modelfile is the one piece that's Spanish-specific: its system prompt and few-shot examples are written for Spanish-to-English glossing.

To study another language, fork ollama/vocab/Modelfile (or replace it in place) with a system prompt and examples for that language, then either:

• run mise run setup (or ankix install --model <name>) to build it under a new Ollama model name, and pass --model/--ollama-model <name> when running kindle/youtube, or
• rebuild the default ankix model in place if you only need one language at a time.

Development

Requires Go and mise.

mise run build       # builds bin/ankix
mise run test         # go test ./...
mise run vet          # go vet ./...
mise run setup        # creates the Ollama model locally
mise run install      # builds and copies ankix to ~/.local/bin