🍍 Msty Admin MCP

AI-Powered Administration for Msty Studio Desktop

An MCP (Model Context Protocol) server that transforms Claude into an intelligent system administrator for Msty Studio Desktop. Query databases, manage configurations, orchestrate local AI models, and build tiered AI workflows—all through natural conversation.

What is This?

Msty Admin MCP lets you manage your entire Msty Studio installation through Claude Desktop. Instead of clicking through menus or manually editing config files, just ask Claude:

"Show me my Msty personas and suggest improvements"

"Compare my local models on a coding task"

"Run calibration tests to see which model handles reasoning best"

"What's the health status of my Msty installation?"

Claude handles the rest—querying databases, calling APIs, analysing results, and presenting actionable insights.

Use Cases

🔍 1. Database Inspection & Insights

Query your Msty database directly through conversation. Access conversations, personas, prompts, knowledge stacks, and MCP tools without touching SQLite.

"Show me all my Msty personas"
"How many conversations do I have?"
"List my configured MCP tools"

🏥 2. Health Monitoring & Diagnostics

Comprehensive health checks for your Msty installation—database integrity, storage usage, model cache status, and actionable recommendations.

"Check the health of my Msty installation"
"Is Sidecar running?"
"How much storage are my models using?"

⚙️ 3. Configuration Sync Between Claude & Msty

Export MCP tool configurations from Claude Desktop and prepare them for Msty import. Generate personas from templates. Convert your Claude preferences to Msty format.

"Export my Claude Desktop MCP tools"
"Generate an Opus-style persona for Msty"
"Sync my preferences to Msty format"

🤖 4. Local Model Orchestration

Direct integration with Msty's Sidecar API. Chat with local models, compare responses across models, and get hardware-aware recommendations.

"List my available local models"
"Chat with qwen2.5:7b about Python async"
"Which model is best for coding on my hardware?"

📊 5. Performance Analytics

Track tokens per second, latency, and error rates across your local models. Privacy-respecting conversation analytics. Identify usage patterns.

"How fast are my local models?"
"Show performance metrics for the last 30 days"
"Which model has the best success rate?"

🎯 6. Model Calibration & Quality Testing

Test your local models against standardised prompts across categories (reasoning, coding, writing, analysis, creative). Score response quality. Track improvement over time.

"Run calibration tests on my Qwen model"
"Test my models on reasoning tasks"
"Show my calibration history"

🔄 7. Tiered AI Workflow (Claude + Local)

Identify which tasks your local models handle well and which should escalate to Claude. Build efficient hybrid workflows where simple tasks go local and complex tasks go to Claude.

"What tasks should I hand off to Claude?"
"Identify patterns where local models fail"
"Compare Claude vs local on this task"

🔬 8. Database Discovery (Advanced)

Through this MCP, we discovered Msty's internal database structure at:

~/Library/Application Support/MstyStudio/File System/000/t/00/00000000

This SQLite database contains tables for:

• personas - Your configured personas
• tools - MCP tool configurations
• toolConfigs - Tool parameters
• conversationTexts - Chat history
• knowledgeStacks - RAG configurations
• And more...

Note: Direct database manipulation is possible when Msty is closed, but unsupported. Use at your own risk.

Key Features

Available Tools (24 Total)

Phase 1: Installation & Health

Phase 2: Configuration Management

Phase 3: Local Model Integration

Phase 4: Intelligence & Analytics

Phase 5: Calibration & Workflow

Installation

Prerequisites

• macOS (Apple Silicon or Intel)
• Python 3.10+
• Msty Studio Desktop installed
• Msty Sidecar running (for local model features)

Quick Start

# Clone the repository
git clone https://github.com/M-Pineapple/msty-admin-mcp.git
cd msty-admin-mcp

# Create virtual environment
python -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

Claude Desktop Configuration

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "msty-admin": {
      "command": "/absolute/path/to/msty-admin-mcp/.venv/bin/python",
      "args": ["-m", "src.server"],
      "cwd": "/absolute/path/to/msty-admin-mcp"
    }
  }
}

Restart Claude Desktop. You should see "msty-admin" in your available tools.

Environment Variables (Optional)

Customise the MCP behaviour with these environment variables:

Example with custom configuration:

{
  "mcpServers": {
    "msty-admin": {
      "command": "/absolute/path/to/msty-admin-mcp/.venv/bin/python",
      "args": ["-m", "src.server"],
      "cwd": "/absolute/path/to/msty-admin-mcp",
      "env": {
        "MSTY_TIMEOUT": "30"
      }
    }
  }
}

Usage Examples

Check Your Installation

You: What's the status of my Msty installation?

Claude: Your Msty Studio is installed at /Applications/MstyStudio.app 
        (version 2.2.0). Sidecar is running with 3 models available.
        Database is healthy at 45MB. No issues detected.

Query Your Data

You: Show me my Msty personas

Claude: You have 3 personas configured:
        1. "Opus Assistant" - British English, quality focus (temp: 0.7)
        2. "Swift Dev Assistant" - Development focused (temp: 0.3)
        3. "Current Ventures" - Business operations (temp: 0.5)

Compare Local Models

You: Compare my local models on this coding task: 
     "Write a Python function to find duplicates in a list"

Claude: Tested 2 models:
        
        qwen2.5-coder:7b
        ├─ Quality: 0.85 | Speed: 38 tok/s | Latency: 2.1s
        └─ Clean implementation with list comprehension
        
        qwen2.5:7b  
        ├─ Quality: 0.72 | Speed: 42 tok/s | Latency: 1.8s
        └─ Works but less elegant
        
        Winner (balanced): qwen2.5-coder:7b

Run Calibration

You: Run calibration tests on my Qwen model for reasoning tasks

Claude: Calibration Results for qwen2.5:7b (Reasoning)
        
        Test 1: Bat and ball problem
        ├─ Score: 0.82 ✅ PASSED
        └─ Correctly identified $0.05
        
        Test 2: Widget machines problem  
        ├─ Score: 0.78 ✅ PASSED
        └─ Showed step-by-step reasoning
        
        Summary: 2/2 passed, average score 0.80

Architecture

┌─────────────────────────────────────────────────────────┐
│                    Claude Desktop                        │
│                         │                                │
│                    MCP Protocol                          │
│                         │                                │
│              ┌──────────┴──────────┐                    │
│              ▼                     ▼                    │
│    ┌─────────────────┐   ┌─────────────────┐           │
│    │ Msty Admin MCP  │   │  Other MCPs     │           │
│    │   (24 tools)    │   │ (Memory, etc.)  │           │
│    └────────┬────────┘   └─────────────────┘           │
└─────────────┼───────────────────────────────────────────┘
              │
   ┌──────────┴──────────┐
   ▼                     ▼
┌──────────┐      ┌──────────────┐
│  Msty    │      │   Sidecar    │
│ Database │      │  Local AI    │
│ (SQLite) │      │   Service    │
└──────────┘      └──────────────┘
     │                   │
     │            ┌──────┴──────┐
     │            ▼             ▼
     │      ┌──────────┐  ┌──────────┐
     │      │ Qwen 2.5 │  │ Llama 3  │
     │      │   7B     │  │   8B     │
     │      └──────────┘  └──────────┘
     │
     ▼
┌────────────────────────────────────┐
│ ~/Library/Application Support/     │
│ MstyStudio/File System/000/t/00/   │
│ 00000000 (SQLite Database)         │
├────────────────────────────────────┤
│ Tables:                            │
│ • personas                         │
│ • tools                            │
│ • toolConfigs                      │
│ • conversationTexts                │
│ • knowledgeStacks                  │
│ • and more...                      │
└────────────────────────────────────┘

Data Storage

The MCP never writes to Msty's database—it only reads. All metrics and calibration results are stored separately.

Hardware Recommendations

For Basic Use (Inspection, Health Checks)

• Any Mac with Msty installed
• No local models required

For Local Model Features

Performance Expectations (Apple Silicon)

FAQ

General

Q: Do I need Msty Studio Desktop installed?
A: Yes. This MCP is specifically designed to administer Msty Studio. Without it, most tools won't function.

Q: Does this work on Windows or Linux?
A: Currently macOS only. Msty Studio Desktop is a macOS application.

Q: Is my data safe?
A: The MCP only reads from Msty's database—it never writes to it. Metrics and calibration data are stored separately in ~/.msty-admin/. No data is sent externally.

Local Models

Q: Do I need local models installed?
A: For basic features (database queries, health checks), no. For local model features (chat, compare, calibrate), you need Msty Sidecar running with at least one model.

Q: Which local models work best?
A: Use recommend_model with your use case. Generally:

• Coding: qwen2.5-coder (7B or 32B depending on your RAM)
• General: qwen2.5 (7B for speed, 32B for quality)
• Fast responses: gemma3:4b or qwen3:0.6b

Q: What's the Sidecar?
A: Msty Sidecar is the background service that hosts local models. It provides an Ollama-compatible API on port 11964.

Q: Can local models use MCP tools?
A: Yes, but smaller models (7B and below) often struggle with complex tool orchestration. Models 14B+ handle tools much better. For reliable MCP tool usage, consider 32B+ models or stick with Claude for complex workflows.

Calibration

Q: What is calibration?
A: Calibration tests your local models against standardised prompts to measure quality. Categories include reasoning, coding, writing, analysis, and creative tasks.

Q: What's a good calibration score?
A: Scores range 0.0-1.0. Generally:

• 0.8+ = Excellent
• 0.6-0.8 = Good (passes threshold)
• 0.4-0.6 = Fair
• Below 0.4 = Poor

Q: What are handoff triggers?
A: Patterns that indicate a task should be handled by Claude instead of a local model. The MCP learns these from failed calibration tests.

Troubleshooting

Q: Claude doesn't see the msty-admin tools
A: Check your claude_desktop_config.json paths are absolute (not relative). Restart Claude Desktop after changes.

Q: "Sidecar not running" error
A: Start Msty Sidecar from the Msty Studio menu bar icon, or ensure Msty is open.

Q: "Database not found" error
A: Msty stores its database in ~/Library/Application Support/MstyStudio/File System/000/t/00/00000000. Ensure Msty has been launched at least once.

Q: Model comparison takes too long
A: Each model runs sequentially. Limit comparisons to 3-5 models. Larger models (32B+) take longer.

Project Structure

msty-admin-mcp/
├── src/
│   ├── __init__.py
│   ├── server.py           # Main MCP server (24 tools)
│   └── phase4_5_tools.py   # Metrics and calibration utilities
├── tests/
│   └── test_server.py
├── requirements.txt
├── pyproject.toml
├── LICENSE
└── README.md

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

Roadmap

• Windows/Linux support (when Msty supports it)
• Direct persona import via API (pending Msty API)
• Automatic model download recommendations
• Integration with Ollama CLI
• Web UI for metrics dashboard

💖 Support This Project

If Claude Command Runner has helped enhance your development workflow or saved you time with intelligent command execution, consider supporting its development:

Your support helps me:

• Maintain and improve Claude Command Runner with new features
• Keep the project open-source and free for everyone
• Dedicate more time to addressing user requests and bug fixes
• Explore new terminal integrations and command intelligence

Thank you for considering supporting my work! 🙏

License

MIT License - see LICENSE for details.

Acknowledgements

• Msty Studio - The excellent local AI application this MCP administers
• Anthropic - For Claude and the MCP protocol
• Model Context Protocol - The foundation making this possible

Created by Pineapple 🍍

Making local AI administration effortless.