Apple Notes search & connection-discovery

apple-notes is an MCP server for semantic search and connection-discovery across the
user's own Apple Notes — hybrid search, Swanson-ABC bridges, entity threads, and cited
synthesis over everything they've written. Embeddings, search, BM25, clustering, and
bridges run on-device; only synthesis generation calls an LLM (local OR cloud,
the user's choice).

This skill covers (1) the one-time setup you must walk the user through, and (2) which
tool to reach for, since the server exposes many.

First: is the MCP connected?

If apple-notes tools are not available, the server isn't registered yet — do the
Setup below before anything else. If tools exist but a search returns "not indexed"
or empty, run index-notes first (see Ranking caveats).

Setup (walk the user through this — it's the skill's real value)

The server reads Apple Notes' SQLite store directly, so the bun binary needs Full
Disk Access. Steps, in order:

1. Install bun (if absent): curl -fsSL https://bun.sh/install | bash
2. Clone + install deps:

   git clone https://github.com/connerkward/mcp-apple-notes
   cd mcp-apple-notes && bun install

1. Grant Full Disk Access to bun. Open System Settings → Privacy & Security → Full

Disk Access, click +, and add ~/.bun/bin/bun (press Cmd+Shift+G in the file
picker and paste the path). Without this the server cannot read NoteStore.sqlite and
every call fails with a permissions error. (bun install's postinstall tries to open
this pane automatically.)

1. Register the MCP server (pick the user's client):

• Claude Code: claude mcp add apple-notes -- bun /absolute/path/to/mcp-apple-notes/index.ts --stdio
• Claude Desktop: add to claude_desktop_config.json:

     { "mcpServers": { "apple-notes": {
         "command": "/Users/<you>/.bun/bin/bun",
         "args": ["/Users/<you>/mcp-apple-notes/index.ts", "--stdio"] } } }

• As a Claude Code plugin (bundles this skill too): /plugin marketplace add connerkward/ckw-skills then /plugin install apple-notes@connerkward.

1. Restart the client, then tell the user to ask "Index my Apple Notes" (or call

index-notes). First index of ~1,800 notes takes a few seconds.

Tool map — which tool for which job

| Tool | Use when |
|------|----------|
| index-notes | First run, or to force a rebuild. Background job with live progress. |
| search-notes | Default search. Hybrid semantic + BM25, re-ranked. Optional folder, modifiedAfter, modifiedBefore. "What did I write about X." |
| find-notes | Exact substring match (like the Apple Notes search box). Use when the user wants a literal string, not meaning. Optional folder, date range. |
| get-note | Fetch one full note by title (fuzzy fallback). |
| list-notes | Notes by recency. Optional folder, date range, limit. |
| list-folders | All folders + note counts. |
| list-tags / search-by-tag | #hashtag inventory / notes carrying a given tag. |
| related-notes | Notes related to a given one via shared tags, [[wikilinks]], and vector similarity. "Show me related notes." |
| bridge-notes | Swanson-ABC bridges — non-obvious connections: pairs (A, C) not directly similar but both strongly tied to a shared intermediary B. "Find non-obvious connections across my notes." Optional folder, limit. No LLM. |
| feed | Ranked evidence-first connection stream (bridges + abstraction pairs + entity threads) as JSON. Optional limit. |
| entity-notes / list-entities | "Where else do I talk about Mercedes?" Entity chips → notes by mention weight. Needs the optional entity graph db (~/.mcp-apple-notes/layered_graph.db); if absent these report how to generate it. |
| get-tables | Pull pipe/tab tables out of a note. |
| create-note / update-note | Create or edit a note. |
| check-changes | Did notes change since last index? (does not trigger re-index) |
| index-health | Sync status, last-indexed time, note count. Run this if results seem stale. |

For "synthesize what I think about X" the synthesis lives in the web app endpoint
(GET /api/synthesize?q= at http://localhost:3741/ when run with bun index.ts),
which writes a grounded answer with inline [n] citations back to source notes.

Ranking caveats (state these when results look off)

• Index before the first search. No index → empty/garbage results; run index-notes.
• Auto re-index: each search does ~1ms change detection and kicks ONE background

incremental index if notes changed — search returns immediately from the current index
and catches up when the job lands. If a just-edited note is missing, it's the catch-up
lag; re-run the search.

• Score: score = RRF(vector, BM25) × title_boost × recency_factor.
• Temporal queries (recent, latest, today) auto-shift to a 1-day recency

half-life at 70% weight; normal queries keep relevance primary (90-day half-life, 10%).

• Synthesis is the only cloud-capable part. It needs an LLM: local via LM Studio /

Ollama (SYNTH_BASE_URL=http://localhost:1234/v1 SYNTH_MODEL= OPENAI_API_KEY=local,
notes stay on-device) or real OpenAI (funded OPENAI_API_KEY, defaults to gpt-4o-mini).
Everything else — embeddings, search, BM25, clustering, bridges, entities — is on-device.

Credits

Fork of RafalWilinski/mcp-apple-notes;
this fork reads SQLite + protobuf directly and adds bridges, entities, feed, and synthesis.