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

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=<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.