`, same iOS purple bar). Mail may still generate its own HTML alternative from the native compose. The default is plain text.\n\n**Compose is not quoted.** `mail_send` / `mail_draft` (plain and html) call `make new outgoing message` **without** AppleScript `content:` so `newMessage` is a real Mail object, then paste the body into the compose window (exact subject title, role `AXWebArea`, proven focus, Cmd-A + clipboard). Mail's `mailto` command was shipped in 2.0.4 and **fails on Mini/MacBook Mail**: it does not return an outgoing message (`newMessage` undefined, AppleScript **-2753**). On current Mail (Ventura+, FB11734014) `content` / `html content` store the body as a citation: every plain-text line prefixed with `>`, plus a `multipart/alternative` HTML part wrapped in ``. Desktop Mail often hides the bar with inline styles; iOS Mail paints the whole body purple with a left quote bar — even when the subject is not `Re:`/`Fwd:` and there is no `In-Reply-To`. Reply and forward still quote the original, which is expected (those paths still prepend via `content`; fixing reply/forward cite-wrap is out of scope for 2.1.0).\n\nKeystrokes use System Events, so **node needs Accessibility** (Privacy & Security → Accessibility) **and** Automation → Mail **and** Automation → System Events. Run `apple-tools-mcp permissions` after first install or upgrade so **node → System Events** pops **Allow** in the same sitting as Mail / Messages / Contacts / Calendar. An Accessibility / System Events deny is not a Mail Automation deny (`-1743` / `-10004`).\n\n**Compose body focus (2.0.8).** After `make new outgoing message`, To and Subject are AppleScript properties (never paste). Cmd-V used to run only after an AX window walker left the header. Quote-prefix Sent prove and Sent/Outbox verify stay gated on fail-closed body/header checks — a MacBook 2.0.7 live `mail_send` with Accessibility allowed still hit `BODY_PASTE_MISDIRECTED` when Cmd-V landed in To/Subject (`ATP-207-SENT-PROVE-20260921-084048`).\n\n**macOS 26 focus compile (2.0.9).** Dual-host Sent prove on `@2.0.8` never reached paste: `atmSafeToPaste` used bare `focused UI element`, which System Events 26.x rejects at compile.\n\n**AppleScript reserved identifiers (2.0.10).** Dual-host Sent prove on `@2.0.9` compiled past the focused-UI fix, then failed in `atmFocusMailBody`: `set {tw, th} to size of atmElem` → **-2741** Expected expression but found `th` (AppleScript reserved ordinal).\n\n**Keystroke compose (2.1.0).** macOS 26 System Events reports `windows of process Mail = 0`, so the AX window walker cannot focus a tall `AXWebArea`. Fresh compose **Tabs into the body** (max Tab×6 as proven on Mini) and **types the body with Return between paragraphs**. Clipboard Cmd-V is only a fallback if a *readable* AX value is still missing the body needle. Fail-closed: missing AX body focus or changed To/Cc/Bcc/subject is `BODY_FOCUS_FAILED` / `BODY_PASTE_MISDIRECTED`; the outgoing message is deleted and nothing is sent. Apple events are wrapped in `with timeout`; hung `osascript` is **SIGKILL**'d so a wedge fails fast instead of sitting ~60s. If the Mini write-bridge / indexer daemon cannot drive System Events, `mail_send` returns `unsupported: true` so the stdio client falls back **in-process** — it does **not** hang with `MAIL_SEND_TIMEOUT_GUIDANCE` for a message that never sent.\n\n**AX body-focus oracle (2.1.1).** Dual-host Mail 16 / macOS 26 showed that Mail AppleScript `content of` the outgoing message is **not** a live oracle for UI-typed/pasted text: `set content` reads back, but after Tab + keystroke/paste into the UI body `content of` stays empty (`typedSeen=false`, bodyLen 0) even when typing still produces a plain Sent `.emlx`. `@2.1.0` calibrated Tab against that property and fail-closed with `BODY_FOCUS_FAILED`. 2.1.1 Tabs until the focused AX role is a body field (`AXWebArea` / non-header `AXTextArea` via `AXFocusedUIElement` / `first UI element whose focused is true` — never `focused UI element`, never `windows of process Mail`). Cmd-V runs only when that AX **value** is readable and still missing the needle. An empty AX value is not a miss (Mail 16 WebArea); header checks still fail-closed. Ship prove that the body landed remains the Sent `.emlx` after send (2.0.7 To+subject verify). Do not “fix Tab count.”\n\n**Compose body hit-test (2.1.3).** Mini host: `AXFocusedUIElement` role coerce fails (**-1700**); `windows of process Mail` is empty; a deep `UI elements` tree walk (the `app_ax_find` class of walk) **truncates after ~23 header/toolbar nodes** and never reaches compose `AXWebArea`. Subject-anchored clipboard read-back is refuted on Mini. The proven path: identify the compose window by the **subject title** among **top-level** System Events UI elements, then **coordinate AX hit-test** (`AXUIElementCopyElementAtPosition`) into the body region. The hit must be `AXWebArea` **\"message body\"**; a short `AXTextField` is fail-closed (`BODY_FOCUS_FAILED`). Focus + `AXValue` produced a **plain** Sent `.emlx` (cite=0) — no AppleScript `content` setter, no physical mouse click. Typed Returns / Cmd-V run only if a readable AX value is still missing the needle. System Events permissions probe is unchanged from 2.1.2.\n\n**AXTitle compile (2.1.4).** Mini host: 2.1.3 `mail_send` never reached hit-test because `osacompile` failed **-2741** on bare `value of attribute \"AXTitle\" of atmEl` in `atmTopElemBits` / `atmTopElemExactTitle` (those handlers sit outside the System Events tell that walks top-level UI elements). Mini solution probe (`results-214/SOLUTION-PROBE-REPORT.md`): wrapping those AXTitle reads inside `tell application \"System Events\"` compiles (`osacompile` exit 0). Hit-test + fail-closed `AXTextField` + plain Sent behavior is unchanged from 2.1.3. Live Sent prove is after npm.\n\n**mail_send success is Sent/Outbox verify (2.0.7).** AppleScript `send` returning without throw is not enough. After a real send the tool looks in **Sent** and **Outbox** and reports success only if the message is there. The success text names the delivery state (`mailbox: sent` + `delivery: sent`, or `mailbox: outbox` + `delivery: outbox` while still sending). If verify misses, the tool returns a failure (`isError`) — not success. Hang/timeout recover matches **To + subject** (and Message-ID when compose captured one). It never matches subject alone (short subjects like `test` are unsafe). `from` / account selection is out of scope; Mail's default From is used. `dry_run` / confirm-blocked writes stay `ok: false`, `planned: true`, `delivered: false`.\n\n**Compile prove (2.1.4 on the Mini host):** Unpatched `buildComposeScript` / `buildMailBodyPasteHandler` `osacompile` **-2741** on those bare AXTitle reads; patched AXTitle-inside-System-Events tell `osacompile` exit 0. Report: `results-214/SOLUTION-PROBE-REPORT.md`. Global install untouched. Live Sent is the 2.1.3 ship prove after npm.\n\n**Probe 3 body delivery (2.1.5).** Mini Probe 3 (`atm-p3-deliver.applescript`, classic AppleScript / System Events, not JXA) proved a click-free path: exact subject window title, first `AXWebArea` from `entire contents` of that window, `set focused of` that element to true, then an independent `AXFocusedUIElement` re-read that must match role `AXWebArea` and position/size before any keystroke. Delivery is Cmd-A + clipboard paste. No Mail `content` / `html content`. No `AXValue` write (`AXUIElementSetAttributeValue` on content is a silent no-op; 2.1.4 `atmAxHitFill` returned `OK` when that value was empty, so the keystroke recovery never ran). No coordinate `click at` (Probe 3: a click can land on another Space or app). Focus mismatch is `BODY_FOCUS_UNPROVEN`: the draft is deleted and nothing is sent. The post-paste body needle runs only when the focused element is still that proven body, so a To: recipient chip (`U+FFFC`) is not `BODY_PASTE_MISDIRECTED`. Sent/Outbox verify, `dry_run`, and confirm are unchanged.\n\n**Manual prove (2.1.3 on the Mini host):** After `apple-tools-mcp permissions` reports System Events granted, `mail_send` a short **multiline** plain message and a short `body_format: \"html\"` message whose body looks like ordinary paragraphs (for example `Quick note
`), neither with a `Re:`/`Fwd:` subject. The body must land via title-marker + hit-test (not Tab / focused-role): no `BODY_FOCUS_FAILED` / `BODY_PASTE_MISDIRECTED`. Inspect each Sent `.emlx`: the text/plain part must contain the intended text (newlines preserved) and must not prefix every body line with `>`, and any HTML alternative must not wrap the whole body in ``. Compose must succeed (no `-2753` / undefined `newMessage`). Prove the production path (write-bridge with node → System Events allowed, **or** the documented in-process fallback). That quote-prefix Mini Sent prove is a separate bar from the 2.0.7 verify contract.\n\nEmails are addressed by their RFC822 **Message-ID**. Pass `message_id`, or pass the `file_path` from `mail_search` / `mail_recent` and the server reads the Message-ID out of the `.emlx` headers for you. `mail_archive` moves the message to its account's Archive (or All Mail) mailbox; `mail_trash` moves it to that account's Trash.\n\n**Send timeout vs TCC — check Sent before retrying.** `mail_send` / `mail_reply` / `mail_forward` can hang after Mail has already put the message in Sent. That hang is a **timeout** (`ETIMEDOUT` / `-1712` / AppleEvent timed out), not a TCC deny. The same hang-vs-TCC split applies to **find / reply / open before send** — an Allowed `ETIMEDOUT` there is never `MAIL_TCC_GUIDANCE`. Real Mail Automation denials report **`-1743`**, **`-10004`**, or “not authorized to send Apple events”. After a send hang the tool looks in Sent (and Outbox) and returns **success** if the message is there — never label a delivered send as TCC fail or timeout. Replies match `In-Reply-To` or an exact recent `Re:` + original subject (the fallback compose does not set reply headers). Forwards match the intended recipient plus an exact `Fwd:` subject or the original Message-ID in a forwarded body — not `In-Reply-To`, and not an unrelated `Fwd:` that only shares a To. Compose (`mail_send`) matches the intended **To plus exact subject** (and Message-ID when available), never subject alone. Neither scan treats the original itself as this send. A real send that returns without throw is also verified the same way before success. If Sent-verify misses, the error tells you it was a hang (or a verify-miss after send returned) and to check Sent. **Clients must Sent-check before retrying a timed-out send**; retrying a message that already landed sends a second copy. This is not a silent TCC grant.\n\n### Messages write tool\n\n| Tool | Arguments | Confirm rule |\n|------|-----------|--------------|\n| `messages_send` | `to[]` or `chat_id`, `text`, `attachment_path`, `service` (`auto` \\| `imessage` \\| `sms`) | `confirm` for multiple handles or a group chat |\n\nSupported identifiers:\n\n- **`to`** — phone numbers in E.164 form (`+15551234567`) or Apple ID email addresses. These map to a Messages `participant` on the iMessage (or SMS relay) service.\n- **`chat_id`** — the chat GUID of an existing conversation, for example `iMessage;-;+15551234567` (1:1) or `iMessage;+;chat123456789` (group). The GUID is checked against `~/Library/Messages/chat.db` before anything is sent, so an unknown chat id is refused rather than delivered somewhere unexpected. The same lookup counts participants, which is how group chats are detected for the confirm rule.\n- **`service`** — `auto` (default) tries iMessage and falls back to the SMS relay; `imessage` and `sms` pin the service.\n- **`attachment_path`** — an absolute path to a file that already exists on this Mac. Text and attachment can be sent together.\n\n### Calendar write tools\n\n| Tool | Arguments | Confirm rule |\n|------|-----------|--------------|\n| `calendar_list_calendars` | none | none (read-only helper) |\n| `calendar_add` | `calendar_name` (required), `title` (required), `start` (required), `end`, `all_day`, `location`, `notes`, recurrence args, `alerts_minutes_before[]` | none |\n| `calendar_edit` | `event_id` (required), optional `eventkit_id` from add, plus any of `title`, `start`, `end`, `location`, `notes`, recurrence args, `alerts_minutes_before[]`, `replace_alerts` | none |\n| `calendar_remove` | `event_id` (required), optional `eventkit_id` from add, optional `calendar_name` | **`confirm` required** |\n| `calendar_rsvp` | `event_id` (required), `response` (`accept` \\| `decline` \\| `tentative`), `attendee_email` | none |\n\nEvents are addressed by their **iCalendar UID**, reported as `Event ID` by `calendar_date` and returned by `calendar_add`. Run `calendar_list_calendars` first so new events land on the intended calendar instead of the default one.\n\n`calendar_add`, `calendar_edit`, and `calendar_remove` share the same write-bridge RPC (`{ tool, args }` into launchd-owned node). They do not use a different socket or calendar account. Calendar.app’s dictionary has **`delete` only** — there is no `remove` or `move to trash`.\n\nNon-recurring `calendar_add` **must** create through **EventKit** and print `via: EventKit` plus `eventkit_id`. Mini `cd74071`: EventKit had `eventKitCalendars=1` but `default=[id NSTaggedPointerString]` — JXA `String(title)` prints the ObjC class, not the calendar name, so the title match missed the only writable calendar. Titles and `calendarIdentifier` are `ObjC.unwrap`d; a single writable calendar or `defaultCalendarForNewEvents` is used when the name does not match. There is **no AppleScript fallback** for non-recurring add.\n\n`calendar_edit` and `calendar_remove` take `eventkit_id` from add (`calendarUUID:eventUUID` on Mini). writeOnly EventKit (`status=4`) can create and read ids off the saved `EKEvent`, but **cannot re-query** via `eventWithIdentifier` — Mini `2cdf44d` failed add on that post-save lookup. Add therefore returns ids from the in-memory event. The indexer write-bridge keeps a long-lived EventKit osascript session that caches the `EKEvent` so edit/remove call `saveEvent` / `removeEvent` without a fetch. When `eventkit_id` is present and the session misses, Calendar.app uid lookup is skipped so iCloud cannot hang. `ETIMEDOUT` / `-1712` is **`timeout`**, never TCC. Failures print `osascript kind=… error=… codes=…`. `--apply` fails closed unless add printed `via: EventKit` and `eventkit_id`.\n\n`--apply` prefers an **On My Mac** calendar when EventKit lists one. `--calendar=` still wins. The Mini host’s first writable calendar named **Calendar** (no On My Mac / iCloud label) is fine if EventKit create actually runs.\n\nLike Contacts, calendar writes go through Calendar.app rather than writing `Calendar.sqlitedb` directly. Reads query that database for speed, but edits must go through the app so iCloud sync, invitations, and alarms behave correctly.\n\n**Supported recurrence patterns.** Either pass structured arguments or a raw `recurrence` RRULE:\n\n| Argument | Values |\n|----------|--------|\n| `frequency` | `daily`, `weekly`, `monthly`, `yearly` |\n| `interval` | 1-366 (e.g. `2` with `weekly` = every other week) |\n| `count` | 1-1000 occurrences (cannot be combined with `until`) |\n| `until` | explicit local datetime; emitted as a UTC `UNTIL` |\n| `by_day` | `MO TU WE TH FR SA SU` (weekly patterns) |\n| `recurrence` | raw RRULE instead of the above, e.g. `FREQ=WEEKLY;INTERVAL=1;COUNT=10` |\n\n`{ frequency: \"weekly\", interval: 2, by_day: [\"MO\",\"WE\"], count: 10 }` becomes `FREQ=WEEKLY;INTERVAL=2;BYDAY=MO,WE;COUNT=10`. Anything outside this grammar is refused.\n\n**Alerts.** `alerts_minutes_before` takes up to 5 whole minute values (0 to 40320, i.e. four weeks) and creates display alarms. On `calendar_edit`, supplying alerts replaces the event's existing alarms; `replace_alerts: true` removes them without adding new ones.\n\n**RSVP.** `calendar_rsvp` sets your attendee participation status on the invitation. Some macOS versions refuse to write that property from AppleScript; when that happens the tool says so explicitly (and asks you to answer in Calendar) rather than reporting a silent success.\n\n### Contacts write tools\n\n| Tool | Arguments | Confirm rule |\n|------|-----------|--------------|\n| `contacts_add` | `first_name`, `last_name`, `organization`, `job_title`, `emails[]`, `email_label`, `phones[]`, `phone_label` | none |\n| `contacts_edit` | `contact_id` (required), any of the above, `replace_emails`, `replace_phones` | **`confirm` required** when replacing with an empty list |\n| `contacts_remove` | `contact_id` (required) | **`confirm` required** |\n\nContacts are addressed by their Contacts.app person id (for example `ABCD1234-...:ABPerson`), reported as `Contact ID` by `contacts_search` and `contacts_lookup` and returned by `contacts_add`. At least one of `first_name`, `last_name`, or `organization` is required to create a contact. Writes go through Contacts.app, never by writing the AddressBook database directly (that breaks iCloud sync). Keep **Contacts, Mail, and Messages** running for write reliability; the Contacts write path launches Contacts.app before CRUD if it was quit. Calendar does not need to stay open (EventKit).\n\n### Example write calls\n\n```jsonc\n// Preview first - nothing is sent\n{ \"name\": \"mail_send\", \"arguments\": { \"to\": [\"a@example.com\"], \"subject\": \"Status\", \"body\": \"All good\", \"dry_run\": true } }\n\n// Single recipient: sends on the first call\n{ \"name\": \"mail_send\", \"arguments\": { \"to\": [\"a@example.com\"], \"subject\": \"Status\", \"body\": \"All good\" } }\n\n// Two recipients: blocked until confirmed\n{ \"name\": \"mail_send\", \"arguments\": { \"to\": [\"a@example.com\", \"b@example.com\"], \"subject\": \"Status\", \"body\": \"All good\", \"confirm\": true } }\n\n// Recurring event with an alert\n{ \"name\": \"calendar_add\", \"arguments\": { \"calendar_name\": \"Work\", \"title\": \"Standup\", \"start\": \"2026-09-21 09:00\", \"end\": \"2026-09-21 09:15\", \"frequency\": \"weekly\", \"by_day\": [\"MO\",\"TU\",\"WE\",\"TH\",\"FR\"], \"alerts_minutes_before\": [10] } }\n\n// Delete: preview, then confirm\n{ \"name\": \"calendar_remove\", \"arguments\": { \"event_id\": \"EVT-UID\", \"confirm\": true } }\n```\n\n### Not included\n\nNo Apple **Reminders** tools ship in this package, by design. Notes, FaceTime, and Files automation are also out of scope.\n\n## Example Queries\n\nAsk your MCP client things like:\n\n- \"Find emails from John about the quarterly report\"\n- \"What messages did I get from Mom last week?\"\n- \"When is my next dentist appointment?\"\n- \"Search for emails about the AWS bill from November\"\n- \"Find all calendar events with Zoom links\"\n- \"What's Sarah's phone number?\"\n- \"Show me all communication with David from last month\"\n\n## Privacy & Security\n\n- **Local Processing**: All embeddings are generated locally using Hugging Face Transformers.js\n- **No Cloud Services**: No data is sent to external servers; the write bridge is a local unix socket, never a network port\n- **Reads are read-only**: Search and lookup tools never modify your data. The [write tools](#write-tools-200) are the only ones that change anything, and they are opt-in per call, with `confirm` required for deletes and multi-recipient sends\n- **No credentials**: The server holds no tokens or passwords. It uses the Mail, Messages, Calendar, and Contacts apps you are already signed into\n- **Your Data**: The vector index is stored locally in your home directory\n\n## Troubleshooting\n\n### \"Authorization denied\" errors\n\nEnsure Node.js has Full Disk Access (see Installation step 2).\n\n### A write tool reports that macOS denied the automation\n\nThe message names which process macOS was actually asking about. Work through it in this order:\n\n1. **MacBook / Terminal (`permissions` CLI):** there is **no** always-on indexer. Re-run `apple-tools-mcp permissions` from **Terminal.app**, Allow the **printed `process.execPath`**, and check System Settings → Privacy & Security → **Automation** for that node → Contacts, Calendar, Mail, Messages, and System Events. Do **not** start `apple-tools-indexer`.\n2. **Mini only:** `pgrep -fl apple-tools-indexer` / `~/.apple-tools-mcp/writer.sock`. If the LaunchAgent is the product host, start it so writes run under launchd-owned node (see [Mini ship-gate](#mini-ship-gate-host-setup)).\n3. **Did you approve the Automation prompts for `node`?** Re-run `apple-tools-mcp permissions` with the product `node` (it prints `process.execPath`). Open System Settings → Privacy & Security → **Automation** and confirm **that printed node** is allowed to control **Mail**, **Messages**, Contacts, Calendar, and **System Events**. A prior **Don't Allow** for System Events is a sticky Deny — the Allow dialog will not appear again; turn **System Events** on for that node, then re-run `permissions` (it must not report granted from a soft name check). On Mini the LaunchAgent binary is whatever the plist / `which node` reports (`/Users/petercoates/.local/node/bin/node` is a Mini *example* only). MacBook Claude nvm is `/Users/petercoates/.nvm/versions/node/v22.21.1/bin/node`, not Homebrew. Watch the host — Allow **`node`**, not any other app. Contacts or Calendar being allowed does **not** grant Mail. A `mail_send` hang is **Automation denied**, not “Mail.app could not be reached”; `dry_run` never talks to Mail so it cannot detect this. Do **not** approve the MCP client / host app that launched a short-lived stdio server, and do **not** try to add `node` via **+** in the Contacts or Calendars privacy lists — those panes often have no Add button. `tccutil reset AppleEvents` re-arms the Automation prompt so you can Allow **`node`** again. Full Disk Access is a separate read grant; npm Trusted Publisher / tokens are unrelated.\n4. **Mini: is the daemon's node binary the one with Full Disk Access?** LaunchAgents do not inherit your shell `PATH`; confirm the plist points at the same path `which node` reports.\n5. **Is it actually the write path?** Contacts or Calendar *reads* failing with `EPERM` is a Full Disk Access / attribution problem, not the entitlement gap — fix FDA for the responsible process. Contacts or Calendar *CRUD* failing with no prompt under Claude Desktop is the [documented host limitation](#reads-and-writes-are-different-mechanisms): Claude.app carries neither the addressbook nor the calendars entitlement. On Mini, start the daemon so the write goes through the bridge. On MacBook, use Terminal.app + the printed node.\n6. **\"Contacts.app / Calendar.app could not be reached\"** with the app clearly installed is an **Automation / responsible-process** failure, not a missing app — macOS reports a refused Apple event as `-1728` / \"can't get application\". On MacBook / Terminal, Allow the printed `process.execPath`. On Mini, the write bridge / LaunchAgent may apply. A Contacts `-600` / “application isn't running” is a **cold launch**, not Automation — keep Contacts, Mail, and Messages running; Calendar does not need to stay open.\n7. **`calendar_remove` FAIL after add/edit PASS** is a delete-path failure, not missing Calendar Automation. Same write-bridge RPC. Read the `osascript kind=/error=` line: `kind=timeout` + `ETIMEDOUT` / `-1712` is an iCloud/CalDAV hang or Calendar confirmation dialog — **not** TCC. `EVENTKIT_NOT_FOUND status=4` is writeOnly EventKit failing to see an AppleScript-created event (add must go through EventKit). `-1743` / `-10004` is a real Automation deny. Paste that line if it still FAILs.\n8. **Prove the Mini host itself works** with `node scripts/smoke-writes.js --apply` on the Node host with the LaunchAgent running; it routes through the bridge and separates the read and write mechanisms for you.\n\n### A write returned \"CONFIRMATION REQUIRED\"\n\nThat is the safety gate, not a failure. Deletes and multi-recipient sends need `confirm: true`; the message states exactly what would have happened. Use `dry_run: true` for a preview.\n\n### Empty search results\n\n1. Check that the index was built: `ls ~/.apple-tools-mcp/vector-index/`\n2. Rebuild the index if needed: `npm run build-index`\n\n### Server not appearing in the MCP client\n\n1. Verify your client config is valid (JSON files must be valid JSON)\n2. Restart the MCP client completely (for Claude Desktop: Cmd+Q, then reopen)\n3. Check the client's MCP logs for errors\n\n### Force rebuild the index\n\nIf the index becomes corrupted or out of sync:\n\n```bash\n# If the Mini indexer LaunchAgent is running, unload it first\n# launchctl unload ~/Library/LaunchAgents/com.apple-tools-mcp.indexer.plist\n\n# Remove existing index files\nrm -rf ~/.apple-tools-mcp/vector-index\nrm -f ~/.apple-tools-mcp/index-meta.json\nrm -f ~/.apple-tools-mcp/indexer.lock\n\n# Restart the indexer daemon or your MCP client to trigger a fresh rebuild\n```\n\n### Monitor indexing progress\n\nWatch the MCP server logs in your client. Log locations vary by client; Claude Desktop example:\n\n```bash\ntail -f ~/Library/Logs/Claude/mcp-server-apple-tools.log\n```\n\n### Audit the index\n\nCheck index health and coverage:\n\n```bash\n# Quick audit\nnpm run audit\n\n# Detailed audit saved to file\nnpm run audit -- --reporter=verbose > audit-report.txt\n```\n\n## Development\n\n```bash\n# Clone the repo\ngit clone https://github.com/sfls1397/Apple-Tools-MCP.git\ncd Apple-Tools-MCP\n\n# Install dependencies\nnpm install\n\n# Install test dependencies\nnpm install -D vitest @vitest/coverage-v8 fast-check\n\n# Run tests\nnpm test\n\n# Run the indexer daemon (owns indexer.lock + vector-index refresh)\nnpm run indexer\n\n# One-shot rebuild (stop the indexer daemon first)\nnpm run build-index\n\n# Run tests with verbose coverage report\nnpx vitest run --coverage --reporter=verbose\n\n# Run audit to check index health\nnpm run audit\n\n# Prove the write path on this host (dry run)\nnpm run smoke:writes\n\n# Real CRUD — the extra -- is required so npm forwards --apply\nnpm run smoke:writes -- --apply\n```\n\n## Contributing\n\nContributions are welcome! Please:\n\n1. Fork the repository\n2. Create a feature branch\n3. Make your changes\n4. Run the tests: `npm test`\n5. Submit a pull request\n\n## License\n\nMIT License - see [LICENSE](LICENSE) for details.\n\n## Acknowledgments\n\n- Built with the [Model Context Protocol SDK](https://github.com/modelcontextprotocol/sdk)\n- Vector search powered by [LanceDB](https://lancedb.com/)\n- Local embeddings via [Hugging Face Transformers.js](https://github.com/huggingface/transformers.js)\n","readmeFilename":"README.md"}