Chat With Your Cards

An AI chat dock inside Anki, with the card you are reviewing and the structure of your collection available as context.

Browse the development-preview gallery.

Chat With Your Cards is built for questions that arise while studying: explain this card, find related material, inspect the surrounding topic, or propose a better note without leaving the reviewer. The agent can search the collection, read notes and cards, summarize deck/tag structure, and prepare changes through a reviewable proposal flow. It can also record native Anki Again reviews on exact cards the learner or an AI grader identifies as wrong, even when those cards are not scheduled today.

Three study moments

Understand the card in front of you. The current reviewer card is supplied as context, so the learner can ask the natural follow-up instead of copying the prompt into a separate chat.

Find the missing prerequisite. The assistant can search the learner’s own collection, explain the conceptual bridge, and suggest a short review path.

Turn a confusion into a proposal. New notes and edits remain pending until the learner reviews the fields, deck, tags, and rationale.

The captures use a synthetic collection created inside a disposable Anki profile. See the user stories for the scenarios and their product intent.

Status

Public preview, v0.1.0. The core chat, collection tools, note proposals, safe application/revert path, and real-Anki GUI tests are working. This first release is intentionally for technically comfortable users while the supported backend set and installation experience broaden.

The supported backend is a locally installed, officially authenticated Claude Code CLI, connected to the add-on through a loopback MCP server. Codex and Pi adapters are planned. CWYC does not accept or store API keys.

Requirements and installation

• Anki 25.09 on macOS or Linux. Windows is experimental in v0.1.0.
• Claude Code 2.1.220 or newer, installed and signed in through its official CLI.

Install Chat With Your Cards from AnkiWeb for automatic updates, or download the exact v0.1.0 .ankiaddon from GitHub. Restart Anki when prompted, then open the dock with Ctrl+J (Cmd+J on macOS). If Claude Code is not available, CWYC opens in a built-in demonstration mode and shows setup instructions; Re-check switches to the real backend without restarting Anki.

Read PRIVACY.md before use. CWYC sends the prompts and collection context needed for a chat through your installed Claude Code CLI. CWYC itself does not collect telemetry.

What it can do

• Keep the current reviewer card prominently in context.
• Search with Anki syntax and inspect notes, cards, decks, tags, and study statistics.
• Build a cached, annotated map of the collection without stuffing the entire collection into every prompt.
• Fetch card images as actual visual inputs rather than passing filenames.
• Propose new notes and edits with deck, tag, note-type, and field constraints.
• Read and change this add-on's own common settings through chat; every change needs approval in the dock, including in trusted collection modes.
• Render the note's real card templates before a proposed change is accepted.
• Show per-field diffs, detect stale edits, and support per-change or whole-session rollback.
• Fail exact future, suspended, buried, and filtered-deck cards through Anki's native scheduler, with a dedicated confirmation/audit chip.
• Search with native Anki syntax at card level without collapsing a matching sibling to its note; results include template, deck, hidden, flag, and scheduling context for safe exact-ID selection.
• Preserve existing hidden state, report it explicitly, and offer a separate “Make available” action that leaves the recorded failure intact.
• Carry neutral Agent Skills for card authoring, curriculum design, safe delivery, and user-owned conventions learned from accepted edits.
• Render Mermaid diagrams and opt-in interactive widgets in a restricted sandbox.

Safety model

Content and deck writes pass through the proposal manager, which validates the operation, displays the change, and applies only what the user accepts. Scheduler grading passes through a separate, narrow grading manager backed by a pinned vendored copy of Safe Collection Operations. It uses Anki's native Grade Now/scheduler operations, exact card IDs, stable note GUID preflight, transactions, postconditions, and idempotent event cursors—never direct scheduling-row writes.

In Propose and Ask-each-read modes, grading waits on its dedicated confirmation chip. Auto-accept and Trusted writes may apply it immediately under their existing session caps; Read-only blocks it. Existing suspension and both forms of burial remain after the failure and are visibly offered for removal. Destructive content actions still create a backup checkpoint first.

Card content is treated as untrusted input. The local MCP server binds to a random loopback port and requires a per-session token; inheritance of unrelated user MCP servers is off by default. Sandboxed widgets have an opaque origin, a no-network content-security policy, and no access to the add-on or collection.

Read SAFETY.md for the threat model and COMPLIANCE.md for the distribution and provider-boundary review.

Development

The add-on itself is pure Python and ships without an npm build step. The React/assistant-ui source under ui/ is compiled into the checked-in web bundle used by Anki.

# Lint, type-check, and run unit tests without Anki
make check

# Run the real add-on in disposable Anki under Docker/Xvfb
make test-gui-smoke-docker

# Fast frontend iteration against the scripted replayer
cd ui && npm run dev

The GUI smoke test uses anki-addon-workbench to install the add-on into a disposable profile, drive the real send and proposal paths, manipulate real test notes, and capture light/dark screenshots. It never opens or changes the user's normal Anki profile.

To recreate all three public user-story captures after the complete smoke suite:

CWYC_PUBLIC_SCREENSHOT=1 uv run --group dev \
  anki-workbench smoke --timeout 120 \
  --screenshot docs/images/chat-with-your-cards.png

See tests/gui_smoke/README.md for platform details.

Repository map

The .ankiaddon artifact is built deterministically from the tracked release configuration in pyproject.toml.

License

MIT. See LICENSE.

Support continued development: ritornello.dev/support.