An MCP server that gives AI assistants access to your Apple Notes, Reminders, and Contacts — with optional BERT-powered semantic search.

All tools run locally via macOS-native APIs (AppleScript, EventKit, Contacts framework). No data leaves your Mac. No API keys required.

New to MCP? Check out the FAQ for answers to common questions about what this is, whether it works with your setup, and how your data stays private.

	Apple Notes & Semantic Search — Connect your AI to Apple Notes with BERT-powered semantic search
	Apple Contacts — Search, create, and manage contacts from your AI assistant
	Apple Reminders — Full CRUD for reminders and lists via any MCP client
	Access Control — Configure exactly which data your AI can access

brew tap bjenkinsgit/tap
brew install psyxe-mcp

This installs everything — binary, Swift helpers, FFmpeg, and the BERT model. No compilation required.

git clone https://github.com/bjenkinsgit/psyxe-mcp.git
cd psyxe-mcp
./build.sh

The build script handles everything automatically:

• Installs Homebrew, Rust, FFmpeg, and pkg-config if missing
• Builds the MCP server binary (Rust)
• Builds Swift helpers for Reminders and Contacts
• Copies helpers next to the binary
• Pre-downloads the BERT model (~90MB) so first search is instant

Build without semantic search (skips FFmpeg and BERT):

./build.sh --no-memvid

Requirements: macOS 12+ (Monterey or later). Xcode Command Line Tools will be prompted if not installed.

The binary and helpers are in target/release/. Use the full path when configuring your MCP client.

Two shortcuts enable linking Reminders to file artifacts:

./install-shortcuts.sh

This opens each shortcut in Shortcuts.app for you to approve.

If you installed via Homebrew, the command is just psyxe-mcp (it's in your PATH). If you built from source, use the full path: /Users/yourname/src/psyxe-mcp/target/release/psyxe-mcp.

claude mcp add psyxe -- psyxe-mcp

Or edit ~/.claude/claude_mcp_config.json:

{
  "mcpServers": {
    "psyxe": {
      "command": "psyxe-mcp"
    }
  }
}

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "psyxe": {
      "command": "psyxe-mcp"
    }
  }
}

Open Settings → MCP Servers → Add new server:

{
  "psyxe": {
    "command": "psyxe-mcp"
  }
}

Edit ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "psyxe": {
      "command": "psyxe-mcp"
    }
  }
}

Edit ~/.codex/config.toml:

[mcp_servers.psyxe]
command = "psyxe-mcp"

Note: If you built from source instead of using Homebrew, replace psyxe-mcp with the full path to the binary (e.g., /Users/yourname/src/psyxe-mcp/target/release/psyxe-mcp).

By default, the MCP server has full access to all your Notes, Reminders, Contacts, and files. To restrict what your AI can see, use the built-in access control CLI to create and manage ~/.psyxe/access.toml.

No manual file editing is needed — the CLI creates the file with secure permissions (owner-only read/write) on first use.

# 1. See what's available
psyxe-mcp access discover reminders
psyxe-mcp access discover notes

# 2. Grant access to only what the AI should see
psyxe-mcp access grant reminders "Work"
psyxe-mcp access grant notes "Projects"

# 3. Verify your restrictions
psyxe-mcp access list

Once any rule is set for a category, only explicitly granted resources are accessible — everything else in that category is denied.

# See your reminder lists
psyxe-mcp access discover reminders

# See your contact groups
psyxe-mcp access discover contacts

# See your note folders
psyxe-mcp access discover notes

# See common file locations
psyxe-mcp access discover files

# Only allow access to specific reminder lists
psyxe-mcp access grant reminders "Work"
psyxe-mcp access grant reminders "Shopping" --rw    # read-write

# Only allow access to a specific contact group
psyxe-mcp access grant contacts "iCloud"

# Only allow access to specific note folders
psyxe-mcp access grant notes "Projects"

# Grant file access to a folder
psyxe-mcp access grant files "/Users/you/Documents" --rw

# Revoke access
psyxe-mcp access revoke reminders "Shopping"

# See current restrictions
psyxe-mcp access list

# Remove all restrictions (restore full access)
psyxe-mcp access reset

Access rules are stored in ~/.psyxe/access.toml with owner-only permissions (chmod 600). The server refuses to load the config if it is group- or world-readable, preventing other processes from tampering with access rights.

When built with the memvid feature (enabled by default), the server includes BERT-powered semantic search for Apple Notes. This uses memvid-rs to encode your notes into a searchable vector index.

The first time you (or your AI assistant) run a semantic search, the server will build an index of all your notes. This takes a few minutes depending on how many notes you have. Subsequent searches are instant.

# Or ask your AI assistant: "search my notes for machine learning concepts"
# It will automatically build the index on first use.

1. All notes are fetched from Notes.app
2. Each note is chunked and encoded with a BERT model (384-dimensional embeddings)
3. Chunks are stored as QR codes in a ProRes video file (compact, durable archive)
4. A vector index (HNSW) enables instant semantic similarity search
5. The index auto-detects when notes change and prompts for rebuild

We've included sample notes designed to showcase semantic search:

# Load 10 sample notes into Apple Notes
./examples/load-sample-notes.sh

# Then ask your AI assistant to rebuild the index and try queries like:
#   "retirement savings"  → finds Tax Strategy (never mentions "retirement")
#   "Italian cooking"     → finds Carbonara recipe (never says "Italian")

See examples/sample-notes.md for the full list of demo queries.

The default model (sentence-transformers/all-MiniLM-L6-v2, 384 dimensions) balances speed and quality. You can swap in any HuggingFace BERT-family sentence-transformer model.

Via environment variable:

MEMVID_MODEL_NAME=BAAI/bge-small-en-v1.5 target/release/psyxe-mcp warmup

Via config file — create memvid_config.toml in the repo root or next to the binary:

[ml]
model_name = "BAAI/bge-small-en-v1.5"

After changing models, rebuild the index (ask your AI assistant or run warmup again).

Popular alternatives:

For instruction-tuned models (like BGE), add query/document prefixes:

[ml]
model_name = "BAAI/bge-small-en-v1.5"
embedding_query_prefix = "Represent this sentence for searching relevant passages: "
embedding_document_prefix = ""

Use any OpenAI-compatible embedding endpoint instead of local BERT:

[ml]
embedding_provider = "remote"

Then set the endpoint via environment variables:

export EMBEDDING_API_URL="http://localhost:11434/v1/embeddings"  # Ollama
export EMBEDDING_API_MODEL="nomic-embed-text"

Works with OpenAI, Ollama, vLLM, LM Studio, or any OpenAI-compatible endpoint.

Build without memvid to skip the FFmpeg/BERT dependency entirely (no brew install needed):

cargo build --release --no-default-features

Notes tools still work — they fall back to AppleScript-based text search. All other tools (Reminders, Contacts, Files) are unaffected.

On first use, macOS will prompt you to grant permission for:

• Notes — "osascript" wants to access Notes
• Reminders — "reminders-helper" wants to access Reminders
• Contacts — "contacts-helper" wants to access Contacts

Approve these in the dialog that appears. You can review/revoke them later in System Settings → Privacy & Security.

┌─────────────────┐     stdio (JSON-RPC)     ┌──────────────┐
│  Claude Code /   │ ◄─────────────────────► │  psyxe-mcp   │
│  Claude Desktop  │                          │  (MCP server) │
└─────────────────┘                          └──────┬───────┘
                                                     │
                                    ┌────────────────┼────────────────┐
                                    ▼                ▼                ▼
                             ┌────────────┐  ┌─────────────┐  ┌───────────┐
                             │ AppleScript │  │ Swift Helper│  │  memvid   │
                             │ (Notes)     │  │ (EventKit,  │  │ (BERT +   │
                             │             │  │  Contacts)  │  │  ProRes)  │
                             └──────┬──────┘  └──────┬──────┘  └─────┬─────┘
                                    ▼                ▼               ▼
                             ┌────────────┐  ┌─────────────┐  ┌───────────┐
                             │  Notes.app  │  │   EventKit  │  │ NoteStore │
                             │             │  │   Contacts  │  │  SQLite   │
                             └────────────┘  └─────────────┘  └───────────┘

The MCP server is a thin stdio bridge. All the real work happens in psyxe-mcp-core, the open-source library that provides direct access to macOS-native APIs.

Place a memvid_config.toml in the working directory or next to the binary:

[chunking]
chunk_size = 700
overlap = 100

[ml]
device = "metal"    # auto | cpu | cuda | metal

[qr]
error_correction = "low"
version = 40

[video]
codec = "prores_ks"
prores_profile = "proxy"
library_log_level = "error"

See memvid-rs for all configuration options.

"osascript is not allowed to send keystrokes" Grant Accessibility permission: System Settings → Privacy & Security → Accessibility

"reminders-helper" wants to access your Reminders Click Allow. If you previously denied, re-enable in System Settings → Privacy & Security → Reminders.

Semantic search is slow on first run The BERT model downloads on first use (~90MB). Subsequent runs use the cached model. Index building speed depends on note count — Metal GPU acceleration helps significantly on Apple Silicon.

Notes search returns stale results The server monitors for changes and will prompt your AI assistant to rebuild the index. You can also force it: ask your assistant to "rebuild the notes index".

Apache 2.0 — see LICENSE.

This project uses FFmpeg at runtime for ProRes video encoding only (LGPL codec). No GPL-licensed codecs (x264, x265, etc.) are used.

Built on psyxe-mcp-core, powered by memvid-rs for semantic search.