Connecting AI tools¶
The MCP server speaks Streamable HTTP at /mcp, with a bearer token:
anki-stack endpoint --show-token
| On this machine | http://127.0.0.1:7012/mcp |
| With HTTPS | https://<domain>/mcp |
| Header | Authorization: Bearer <MCP_BEARER_TOKEN> |
Tools¶
| Tool | Actions |
|---|---|
anki_notes |
add (a missing deck is created), bulk_add (up to 50), find (Anki's search syntax), get_info, update (fields and tags), bulk_update (up to 50), delete (needs confirm: true) |
anki_metadata |
list_decks, create_deck, list_models (note types), model_fields, list_tags |
anki_sync |
Sync now (it also happens on its own after every change) |
A duplicate (same first field, same note type, same deck) is refused with the reason, unless allow_duplicate is set. Errors from Anki come back as the tool's error message.
Clients¶
Claude Code:
claude mcp add --transport http anki https://<domain>/mcp --header "Authorization: Bearer <token>"
Other clients (Claude Desktop through a bridge, Cursor, an MCP gateway): an HTTP MCP server at the URL above, with the header.
The desktop's AnkiConnect instead¶
With the desktop running the real AnkiConnect add-on, the MCP server can use it instead of the worker: MCP_ANKI_URL=http://anki-desktop:8765 in .env, then anki-stack restart mcp. The tools are the same. The worker is the better default: it doesn't depend on a desktop app staying open.