viyv-mcp 2.0.1

A wrapper library for MCP (Model Context Protocol) + Starlette

viyv_mcp

viyv_mcp is a production-ready Python wrapper around FastMCP and Starlette that simplifies creating MCP (Model Context Protocol) servers with minimal boilerplate.

🚀 Quick Start

# Install the package
pip install viyv_mcp

# Create a new MCP server project
create-viyv-mcp new my_mcp_server

# Navigate to the project and install dependencies
cd my_mcp_server
uv sync

# Run the server
uv run python main.py

Your MCP server is now running at http://localhost:8000 🎉

✨ Key Features

🛠️ Simple Tool Creation

from viyv_mcp import tool

@tool(description="Add two numbers")
def add(a: int, b: int) -> int:
    return a + b

🌉 External MCP Server Bridge

// app/mcp_server_configs/filesystem.json
{
  "command": "npx",
  "args": ["@modelcontextprotocol/server-filesystem", "/workspace"],
  "env": {
    "API_KEY": "$API_KEY"  // Environment variable interpolation
  },
  "cwd": "/path/to/working/dir",  // Optional
  "tags": ["filesystem", "io"]     // Optional: for filtering
}

🚀 Production-Ready Multi-Worker Support (New in v0.1.10)

# Enable stateless HTTP mode for multi-worker deployments
STATELESS_HTTP=true uv run python main.py

# Deploy with Gunicorn (recommended for production)
uv pip install gunicorn
STATELESS_HTTP=true uv run gunicorn main:app \
  -w 4 -k uvicorn.workers.UvicornWorker -b 0.0.0.0:8000

🔒 JWT Authentication & Access Control

# Tools with security metadata
@tool(
    description="Query salary",
    namespace="hr",                 # Visible only to hr agents
    security_level="confidential",  # Requires confidential+ clearance
)
def query_salary(employee_id: str) -> str:
    return f"Salary: $100,000"

# Generate JWT for an agent
python -m viyv_mcp generate-jwt \
  --sub hr-agent --clearance confidential --namespace hr \
  --trust common --expires 24h --secret "$SECRET"

• Namespace: Controls tool visibility (tools/list filtering)
• Security Level: Controls tool executability (tools/call clearance check)
• Modes: bypass (dev), authenticated (JWT), deny_all (default safe)
• Both transports: Works for stdio (.mcp.json) and HTTP

🔗 Built-in Integrations

• Custom FastAPI Endpoints: Mount additional APIs with @entry

📦 Installation

# Basic installation
pip install viyv_mcp

# With optional security config (YAML support)
pip install "viyv_mcp[security]"

📁 Project Structure

When you create a new project with create-viyv-mcp new my_project:

my_project/
├── main.py                # Server entry point
├── pyproject.toml         # Dependencies (managed by uv)
├── Dockerfile             # Production-ready container
├── .env                   # Environment variables
└── app/
    ├── config.py          # Configuration management
    ├── tools/             # MCP tools (@tool decorator)
    ├── resources/         # MCP resources (@resource decorator)
    ├── prompts/           # MCP prompts (@prompt decorator)
    ├── agents/            # AI agents (@agent decorator)
    ├── entries/           # Custom HTTP endpoints (@entry decorator)
    └── mcp_server_configs/ # External MCP server configurations

💻 Advanced Usage Examples

Creating Resources with URI Templates

from viyv_mcp import resource

def register(mcp):
    @resource("database://{table}/{id}")
    def get_record(table: str, id: str) -> dict:
        """Fetch a database record by table and ID"""
        # Your database logic here
        return {"table": table, "id": id, "data": "..."}

Prompts with Parameters

from viyv_mcp import prompt
from typing import Annotated
from pydantic import Field

def register(mcp):
    @prompt("code_review")
    def code_review_prompt(
        language: Annotated[str, Field(description="Programming language")],
        code: Annotated[str, Field(description="Code to review")]
    ) -> str:
        return f"""
        Please review this {language} code:
        ```{language}
        {code}
        ```
        Focus on: performance, security, and best practices.
        """

📂 Tool Grouping (New in v0.1.13)

Organize tools into groups for better discoverability and UI presentation:

from viyv_mcp import tool

def register(mcp):
    @tool(
        description="Add two numbers",
        group="計算ツール",  # Group name
        title="加算"         # UI display name (optional)
    )
    def add(a: int, b: int) -> int:
        return a + b

    @tool(
        description="Subtract two numbers",
        group="計算ツール"  # Same group
    )
    def subtract(a: int, b: int) -> int:
        return a - b

    @tool(
        description="Delete a file",
        group="ファイルシステム",
        destructive=True  # Destructive operation hint
    )
    def delete_file(path: str) -> bool:
        import os
        os.remove(path)
        return True

External MCP Server Grouping:

// app/mcp_server_configs/filesystem.json
{
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-filesystem", "/workspace"],
  "group": "ファイルシステム",  // Apply to all tools
  "group_map": {                 // Override per tool (optional)
    "read_file": "ファイル操作/読み込み",
    "write_file": "ファイル操作/書き込み"
  }
}

How it works:

• Group information is stored in _meta.viyv.group (vendor namespace)
• MCP clients can use groups for organized display
• Backward compatible: tools without groups work normally

🔧 Configuration

Environment Variables

# Server Configuration
HOST=0.0.0.0                    # Server host (default: 127.0.0.1)
PORT=8000                        # Server port (default: 8000)
STATELESS_HTTP=true              # Enable stateless mode for multi-worker

# Directory Configuration
BRIDGE_CONFIG_DIR=app/mcp_server_configs  # External MCP configs
STATIC_DIR=static/images                  # Static file serving

Configuration Class

# app/config.py
from viyv_mcp.app.config import Config

class MyConfig(Config):
    # Inherit base configuration

    @staticmethod
    def get_stateless_http():
        """Get stateless HTTP setting from environment"""
        env_val = os.getenv("STATELESS_HTTP", "").lower()
        if env_val in ("true", "1", "yes", "on"):
            return True
        elif env_val in ("false", "0", "no", "off"):
            return False
        return None  # Use FastMCP default

🏗️ Architecture & Advanced Features

ASGI-Level Routing (SSE Streaming Fix)

viyv_mcp implements custom ASGI routing to fix SSE streaming issues:

• Direct /mcp path routing bypasses Starlette middleware
• Ensures proper Server-Sent Events handling
• Compatible with FastMCP's streaming protocol

Dynamic Tool Injection

• Tools are refreshed on every request
• Agents always have access to the latest tools
• Supports runtime tool filtering with tags

External MCP Server Management

• Child process management with stdio communication
• Automatic tool/resource/prompt registration
• Environment variable interpolation in configs
• Tag-based filtering for selective tool inclusion

Production Deployment Features

Stateless HTTP Mode

• No session ID requirements
• Perfect for load-balanced environments
• Enable with STATELESS_HTTP=true

Multi-Worker Deployment

# test_app.py - Create a module for Gunicorn
from viyv_mcp import ViyvMCP
from app.config import Config

stateless_http = Config.get_stateless_http()
app = ViyvMCP("Production Server", stateless_http=stateless_http).get_app()

# Run with Gunicorn
gunicorn test_app:app -w 4 -k uvicorn.workers.UvicornWorker -b 0.0.0.0:8000

Docker Deployment

FROM python:3.10-slim
WORKDIR /app
COPY . .
RUN pip install uv && uv sync
CMD ["uv", "run", "gunicorn", "main:app", "-w", "4", "-k", "uvicorn.workers.UvicornWorker", "-b", "0.0.0.0:8000"]

📊 Performance Optimization

Caching Strategies

• Tools are cached per request
• External MCP connections are persistent
• Static file serving with efficient caching headers

Resource Management

• Automatic cleanup of external MCP processes
• Connection pooling for external services
• Graceful shutdown handling

Monitoring & Debugging

# Enable debug logging
LOG_LEVEL=DEBUG uv run python main.py

# Health check endpoint
curl http://localhost:8000/health

🔍 Troubleshooting

Common Issues

SSE Streaming Not Working

• Ensure no middleware interferes with /mcp path
• Check ASGI routing configuration
• Verify FastMCP version >= 2.12.3

Multi-Worker Startup Failures

• Enable STATELESS_HTTP=true for multi-worker mode
• Use Gunicorn instead of uvicorn's --workers flag
• Check for asyncio event loop conflicts

External MCP Server Issues

# Check external server logs
tail -f logs/external_mcp.log

# Verify command exists
which npx

# Test configuration
cat app/mcp_server_configs/test.json | jq .

Protocol Compatibility

• Use MCP protocol version 2024-11-05
• Pydantic v2 compatibility is patched automatically
• Check mcp_initialize_fix.py for validation patches

📚 Examples

Complete working examples in the example/ directory:

• claude_code_mcp: Claude Code CLI integration
• test: Comprehensive example with all features
  • External MCP servers
  • Custom endpoints

🤝 Contributing

We welcome contributions! Please see our Contributing Guide.

Development Setup

# Clone the repository
git clone https://github.com/BrainFiber/viyv_mcp
cd viyv_mcp

# Install in development mode
pip install -e .

# Run tests
pytest

# Build package
python -m build

# Run example project
cd example/test
uv sync
STATELESS_HTTP=true uv run python main.py

Testing Guidelines

• Add sample implementations in test/ directory
• Test with both session and stateless modes
• Check external MCP server bridging

📄 License

MIT License - see the LICENSE file for details.

👥 Authors

• Hiroki Takezawa - BrainFiber

🙏 Acknowledgments

• Built on FastMCP by jlowin
• Powered by Starlette ASGI framework
• Implements Model Context Protocol specification

📮 Support

• 📧 Email: [e-mail]
• 🐛 Issues: GitHub Issues
• 💬 Discussions: GitHub Discussions
• 📖 Documentation: Wiki

🚦 Roadmap

• WebSocket support for real-time communication
• Built-in authentication/authorization
• Tool versioning and migration support
• Performance profiling dashboard
• Plugin system for custom integrations
• GraphQL endpoint support

📈 Changelog

v0.1.14 (Latest - 2025-10-13)

• 📚 Implementation Examples: Added comprehensive tool grouping examples in example/test
  • 9 internal tools with group organization (Math, Statistics, Web Search, Image Tools)
  • Real-world Playwright MCP server configuration (20 browser automation tools)
• 📖 Enhanced Documentation:
  • GROUPING_IMPLEMENTATION.md - detailed implementation report
  • app/mcp_server_configs/README.md - external MCP server grouping guide
  • Sample configuration files for learning
• ✅ Verified Implementation: All examples tested and working with MCP Inspector
• 🎯 Makes tool grouping feature (v0.1.13) immediately usable with practical examples

v0.1.13

• 📂 Tool Grouping: Organize tools with group parameter in @tool and @agent decorators
• 🏷️ Vendor Namespace: Uses _meta.viyv.group for MCP spec compliance
• 🌉 External MCP Grouping: Support group and group_map in mcp_server_configs/*.json
• ✨ Optional Parameters: Added title and destructive hints
• 🔄 Backward Compatible: Tools without groups work normally
• 📚 Enhanced templates and documentation with grouping examples

v0.1.10

• ✨ Added stateless HTTP support for multi-worker deployments
• 🔧 Improved ASGI-level routing for SSE streaming
• 📦 Updated FastMCP to 2.12.3 for better compatibility
• 🐛 Fixed Pydantic v2 validation issues
• 📚 Enhanced documentation and examples

v0.1.9

• 🌉 External MCP server bridging with tags and filtering
• 🔄 Dynamic tool refresh on every request

---

Made with ❤️ by the viyv_mcp community