AppleMCP

Native macOS 15+ MCP server that gives AI assistants secure, read-only access to your local Apple data and Apple Intelligence features.

AppleMCP consists of a SwiftUI app that bridges macOS privacy-controlled APIs and a stdio MCP server for Claude Desktop, Claude Code, and other MCP clients. Everything runs locally — no cloud, no sync, no data leaves your machine.

What It Does

Source	Access Method	Permission
Mail	Local Envelope Index (SQLite) + `.emlx` body parsing, AppleScript fallback	Full Disk Access or Mail Automation
Calendar	EventKit	Calendar Access
Contacts	Contacts.framework	Contacts Access
Reminders	EventKit	Reminders Access
Notes	Notes.app AppleScript Automation	Automation Permission
Photos	Photos.framework	Photos Access
Voice Memos	Local `CloudRecordings.db` + in-file transcripts, Speech.framework for on-device recognition	Full Disk Access, Speech Recognition (only for `voicememos_transcribe`)
Apple Intelligence	Native APIs (ImagePlayground, Translation, Writing Tools)	None

Quick Start

1. Build

swift build
swift test   # parser tests for the Voice Memos transcript format

2. Run the App

./script/build_and_run.sh

The app opens a macOS window and starts a local HTTP endpoint on 127.0.0.1:47651. Click Permissions on first launch to grant macOS access to Calendar, Contacts, Reminders, Photos, and Notes.

3. Connect Your MCP Client

Claude Desktop

Add to your Claude Desktop MCP config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "applemcp": {
      "command": "/path/to/AppleMCP/.build/debug/M3MCPBridge"
    }
  }
}

Claude Code

claude mcp add applemcp /path/to/AppleMCP/.build/debug/M3MCPBridge

The M3MCP UI app must be running for MCP calls to work. The bridge communicates with the app over a local loopback HTTP connection.

MCP Tools

Data Access

Tool	Description
`mail_search`	Search messages in Apple Mail (subject, sender, date, read status)
`mail_read`	Read full email content by ID (body, recipients, attachments metadata)
`calendar_search`	Search calendar events via EventKit
`contacts_search`	Search contacts / address book
`reminders_search`	Search reminders (incomplete, completed, or all)
`notes_search`	Search Apple Notes by keyword
`notes_read`	Read a single note by ID
`photos_search`	Search Apple Photos library metadata
`photos_albums`	List photo albums with counts
`voicememos_search`	Search voice memos by title, date range, or transcript text
`voicememos_read`	Read one recording including its stored transcript
`voicememos_transcript`	Return a stored transcript as text, timestamped text, or JSON segments
`voicememos_audio`	Return the recording as a local path or base64 audio
`voicememos_transcribe`	Transcribe a recording with on-device speech recognition

Apple Intelligence

Tool	Description
`ai_writing_tools`	Summarize, rewrite, proofread, or change tone of text
`ai_translate`	Translate text using Apple system translation
`ai_image_playground`	Generate images from text descriptions (macOS 15.4+)

System

Tool	Description
`source_status`	List available providers and their states
`permissions_status`	Report macOS permission state for all providers
`permissions_request`	Request all required macOS permissions
`permissions_open_settings`	Open System Settings for permission remediation

Architecture

MCP Client (Claude) <--stdio--> M3MCPBridge <--HTTP 127.0.0.1:47651--> M3MCPApp (SwiftUI)
                                                                            |
                                       EventKit / Contacts / Photos / Mail Index / Voice Memos Store / Speech / AppleScript

M3MCPApp — SwiftUI app with macOS TCC permissions, provides the actual data access
M3MCPBridge — Lightweight stdio MCP server that translates MCP protocol to HTTP calls
M3MCPCore — Shared models and types

Voice Memos

Voice Memos are read straight from the local Core Data store (~/Library/Group Containers/group.com.apple.VoiceMemos.shared/Recordings/CloudRecordings.db) — Voice Memos.app is never driven through AppleEvents. Transcripts have no sidecar file: macOS writes them into a private tsrp atom inside each .m4a, so AppleMCP parses the recording itself.

Open Voice Memos once so macOS creates the store, and grant Full Disk Access if the folder is protected.
On macOS Sequoia and later, opening a memo in Voice Memos makes macOS transcribe it. voicememos_transcript then returns that transcript without any recognition run.
voicememos_transcribe falls back to Speech.framework and prefers on-device recognition, so audio stays on the Mac. It returns an existing transcript first unless you pass prefer_stored: false.
voicememos_search matches titles by default. Pass search_transcripts: true to search spoken content instead; max_candidates bounds how many recordings are opened.

See docs/VOICE_MEMOS.md for the store layout, the transcript format, and troubleshooting.

Privacy

All data stays local. The app uses macOS TCC (Transparency, Consent, and Control) for every data source. No network requests are made except to 127.0.0.1. Mail reads the local SQLite index directly — it never sends emails or modifies any data. Voice Memos are opened read-only, and speech recognition runs on device whenever the locale supports it, so recordings never leave the Mac.

Requirements

macOS 15.0+
Swift 5.9+
Apple Intelligence features require macOS 15.4+ with Apple Silicon

Credits

The Voice Memos support is a native Swift port of jwulff/apple-voice-memo-mcp (MIT). See docs/THIRD_PARTY.md.

License

Apache License 2.0 — see LICENSE.