Apple Agent Orchestrator

Multi-agent orchestration through the Model Context Protocol (MCP) with privacy-first design. Demonstrates how an OS-level agent system can route complex multi-step tasks to domain-specific agents while keeping sensitive data on-device.

Built for Apple's emerging architecture: App Intent Domains + MCP integration + on-device privacy.

graph TD
    User[User Request] --> Orchestrator[Orchestrator Agent]
    Orchestrator --> Router[Intent Router]
    Router --> |messaging| MsgAgent[Messaging Agent]
    Router --> |productivity| ProdAgent[Productivity Agent]
    Router --> |media| MediaAgent[Media Agent]
    Router --> |knowledge| KnowledgeAgent[Knowledge Agent]

    MsgAgent --> MCP1[MCP Server: Messaging]
    ProdAgent --> MCP2[MCP Server: Productivity]
    MediaAgent --> MCP3[MCP Server: Media]
    KnowledgeAgent --> MCP4[MCP Server: Knowledge]

    MCP1 --> Tools1[compose, reply, search_messages]
    MCP2 --> Tools2[create_task, schedule, notes]
    MCP3 --> Tools3[play, search_media, playlist]
    MCP4 --> Tools4[search, summarize, lookup]

    Orchestrator --> Privacy[Privacy Guard]
    Privacy --> |PII detection| Redact[On-Device Redaction]
    Privacy --> |data classification| Classify[Sensitivity Classifier]

    subgraph On-Device Processing
        Orchestrator
        Router
        Privacy
        Redact
        Classify
    end

    subgraph Domain Agents
        MsgAgent
        ProdAgent
        MediaAgent
        KnowledgeAgent
    end

    subgraph MCP Servers
        MCP1
        MCP2
        MCP3
        MCP4
    end

Demo

https://github.com/hashwnath/apple-mcp-agent-orchestrator/releases/download/v1.0/demo.mp4

Live vs Blueprint

Feature	Status
Orchestrator Agent with LangGraph state machine	Live
Intent Router (4 App Intent Domains)	Live
MCP Server with tool registration	Live
MCP Client communication	Live
Privacy Guard (PII detection + redaction)	Live
Data sensitivity classification	Live
Domain agents (messaging, productivity, media, knowledge)	Live
Multi-step task decomposition	Live
Cross-domain dependency resolution	Live
CLI demo interface	Live
25 passing tests	Live
Real iOS app integration	Blueprint
Neural Engine optimization	Blueprint
Core ML model integration	Blueprint
Siri voice interface	Blueprint

Quick Start

# Clone
git clone https://github.com/hashwnath/apple-mcp-agent-orchestrator.git
cd apple-mcp-agent-orchestrator

# Install
pip install -r requirements.txt

# Run demo
python -m src.main --demo

# Interactive mode
python -m src.main

# Single request
python -m src.main --request "Check my unread messages from Sarah and create tasks for the action items"

# Run tests
pytest tests/ -v

How It Works

1. User sends a natural language request

"Check my unread messages from Sarah, summarize them, create a task for the action items, and play some focus music while I work on them."

2. Intent Router classifies into App Intent Domains

• messaging: search_messages (sender=Sarah, unread)
• knowledge: summarize (depends on messaging results)
• productivity: create_task (depends on knowledge extraction)
• media: play_music (mood=focus, independent)

3. Orchestrator executes with dependency resolution

• Independent tasks run in parallel (messaging + media)
• Dependent tasks wait for upstream results (knowledge waits for messages, productivity waits for extracted action items)

4. Privacy Guard enforces at every boundary

• PII detection before any data crosses agent boundaries
• Sensitivity classification (public/internal/confidential/restricted)
• Automatic redaction for confidential+ data
• Audit logging of all privacy decisions

Architecture

App Intent Domains mirror Apple's architecture for agentic app control:

• Messaging - emails, texts, notifications
• Productivity - tasks, calendars, notes, reminders
• Media - music, podcasts, playback control
• Knowledge - search, summarization, information retrieval

MCP Layer provides standardized agent-tool communication:

• Each domain agent has its own MCP server with registered tools
• MCP clients handle request/response lifecycle
• Privacy metadata travels with every MCP message

Privacy Guard enforces Apple's privacy hierarchy:

1. On-device (default) - restricted and confidential data never leaves
2. Private Cloud Compute - internal data may use secure cloud
3. External - only public data, only with consent

About

Built by Hashwanth Sutharapu - SDE at MAQ Software (Microsoft Partner), working on multi-agent orchestration and MCP servers in production.

Relevant experience:

• ThinkNet: 8-agent orchestration system on AWS Bedrock
• Production MCP servers handling 1K+ requests/minute
• Contributor to Microsoft Agent Framework (8K+ stars)
• Computer Use Agent with LangChain AgentExecutor
• Enterprise agents for MCAPS teams at Microsoft

This project demonstrates patterns directly applicable to Apple's App Intents + MCP architecture for agentic AI on iOS/macOS.