Search and discover connections across Apple Notes

This skill drives the apple-notes MCP server for semantic and keyword search and connection-discovery across a user's own Apple Notes. It offers hybrid search, Swanson-ABC bridges, entity threads and cited synthesis over everything the user has written. Embeddings, search, BM25, clustering and bridges run on-device, and only synthesis generation calls an LLM, which can be local or cloud at the user's choice. It runs on macOS.

The skill covers two things: the one-time setup the agent must walk the user through, and which tool to reach for, since the server exposes many.

The tool map:

• index-notes builds or rebuilds the index as a background job with live progress. index-health reports sync status, last-indexed time and note count, and check-changes says whether notes changed since the last index without re-indexing.
• search-notes is the default: hybrid semantic plus BM25, re-ranked, with optional folder and date filters. find-notes does exact substring matching like the Apple Notes search box.
• get-note, list-notes (by recency), list-folders (with note counts), list-tags and search-by-tag retrieve and browse notes. get-tables pulls pipe or tab tables out of a note, and create-note and update-note create or edit notes.
• related-notes finds notes related through shared tags, wikilinks and vector similarity.
• bridge-notes finds Swanson-ABC bridges: pairs (A, C) not directly similar but both strongly tied to a shared intermediary B. It uses no LLM. feed returns a ranked, evidence-first connection stream as JSON.
• entity-notes and list-entities answer questions such as where else a name appears. They need an optional entity graph database, and report how to generate it when it is missing.

Ranking works as score = RRF(vector, BM25) x title_boost x recency_factor. Temporal queries such as recent, latest or today shift to a 1-day recency half-life at 70 percent weight, while normal queries keep relevance primary with a 90-day half-life at 10 percent. Each search does about 1 ms of change detection and starts one background incremental index if notes changed, so a just-edited note may be missing until the job lands.

Use it to find, recall or look up something from the user's notes, to surface non-obvious connections or related notes, to synthesize a position from notes, and to index notes or run tag and folder queries. Do not use it for creating reminders or for other note systems: it does not search Obsidian, Notion, Google Docs or other stores. Setup cannot be completed purely inside a remote shell because the server needs local filesystem permissions. Search quality depends on a fresh index, and entity tools do not work without the optional graph database.

Input is a natural-language request about the user's notes. Output is ranked notes, connection lists or a grounded synthesis with inline numbered citations back to source notes, served through a local web endpoint at port 3741. Synthesis needs an LLM, either local through LM Studio or Ollama, which keeps notes on-device, or OpenAI with a funded key, which defaults to gpt-4o-mini.

The server reads Apple Notes' SQLite store directly, so the bun binary needs Full Disk Access; without it every call fails with a permissions error. Steps: install bun, clone and install the server, grant Full Disk Access to that exact bun path, register the server, restart the client and ask to index the notes. The first index of about 1,800 notes takes a few seconds.

claude mcp add apple-notes -- bun /absolute/path/to/mcp-apple-notes/index.ts --stdio

The server can also be registered in Claude Desktop through claude_desktop_config.json or installed as a Claude Code plugin.

It works through the apple-notes MCP server, a fork of RafalWilinski/mcp-apple-notes that adds bridges, entities, feed and synthesis, with Claude Code and Claude Desktop as clients, and LM Studio, Ollama or OpenAI for synthesis.

Mac users with large Apple Notes collections who want to search by meaning, rediscover forgotten notes and link ideas across them.