Quick Start

Install

pip install -e ".[dev]"

Run the Demo

# Enterprise agent demo (standalone, no Azure credentials needed)
python -m demo.enterprise_agent

# Multi-agent workflow demo
python -m demo.workflow_demo

# Mock MCP server (for testing)
python -m demo.mock_mcp_server

Run Tests

pytest tests/ -v

Use with Agent Framework

import asyncio
from mcp_skills import MCPSkillsProvider, MCPClientConfig

async def main():
    # 1. Configure MCP connection
    config = MCPClientConfig(
        server_uri="http://your-mcp-server:8080",
        server_name="enterprise-mcp",
        pool_size=10,
    )

    # 2. Initialize provider (auto-discovers skills from MCP server)
    provider = MCPSkillsProvider(
        client_config=config,
        skills_dir="./skills",
        auto_discover=True,
    )
    await provider.initialize()

    # 3. See what was discovered
    print(f"Skills: {provider.get_skill_names()}")
    # -> ['database_query', 'document_search', 'api_gateway']

    # 4. Call tools directly
    response = await provider.call_tool(
        "database_query",
        {"sql": "SELECT * FROM customers ORDER BY revenue DESC LIMIT 10"}
    )
    print(f"Result: {response.result}")

    # 5. Or use with Agent Framework agent
    # agent = client.as_agent(context_providers=[provider])

    await provider.shutdown()

asyncio.run(main())

Key Features

Auto-Discovery

Point at any MCP server. Skills are auto-generated with proper SKILL.md frontmatter, following Agent Framework's progressive disclosure model (advertise -> load -> read).

Connection Pooling

Configurable pool size with health checks and automatic reconnection. Designed from production experience running MCP at 1K+ requests/minute.

Middleware Stack

Composable middleware chain with sensible defaults:

• Rate Limiting - Token bucket algorithm (configurable RPM + burst)
• Retry - Exponential backoff with jitter (configurable max retries)
• Telemetry - OpenTelemetry spans with tool name, latency, success/error
• Logging - Request/response logging at configurable levels

from mcp_skills.middleware import MiddlewareStack, RateLimiter, RateLimiterConfig

# Custom middleware stack
stack = MiddlewareStack()
stack.add(TelemetryMiddleware("my-service"))
stack.add(RateLimiter(RateLimiterConfig(requests_per_minute=120, burst_size=20)))
stack.add(RetryMiddleware(max_retries=5))

provider = MCPSkillsProvider(middleware=stack)

Progressive Disclosure

Follows Agent Framework's three-stage skill model:

1. Advertise (~100 tokens) - Name + description in system prompt
2. Load (<5K tokens) - Full SKILL.md instructions via load_skill
3. Read - Reference docs via read_skill_resource

Project Structure

microsoft-cape-mcp-skills/
├── src/mcp_skills/
│   ├── client.py         # MCP client pool with connection management
│   ├── discovery.py      # Auto-discovery and SKILL.md generation
│   ├── middleware.py      # Rate limiting, retry, telemetry
│   ├── provider.py       # Agent Framework SkillsProvider integration
│   └── telemetry.py      # OpenTelemetry configuration
├── skills/               # Pre-built enterprise skills
│   ├── database_query/   # SQL query skill with safety validation
│   ├── document_search/  # Document repository search skill
│   └── api_gateway/      # Enterprise API gateway skill
├── demo/
│   ├── enterprise_agent.py   # Full agent demo
│   ├── workflow_demo.py      # Multi-agent workflow demo
│   └── mock_mcp_server.py    # Local MCP server for testing
└── tests/                # Comprehensive test suite

Why This Exists

Microsoft Agent Framework v1.0 introduced Agent Skills and MCP support. The roadmap calls for expanding the skills ecosystem. This package demonstrates how to bridge MCP servers into the Agent Framework skill system with enterprise-grade patterns.

Built from real-world experience:

• Contributing to Microsoft Agent Framework (9.3K stars)
• Running production MCP servers at 1K+ req/min
• Building enterprise agents for Microsoft MCAPS teams
• Multi-agent orchestration with ThinkNet (8-agent system on AWS Bedrock)

Precedents

• AgentEval - Community-built .NET evaluation toolkit for Agent Framework
• Microsoft acqui-hired the Cove AI team to accelerate CoreAI agent development
• Agent Framework v1.0 explicitly calls for community contributions to expand the skills ecosystem

About

Built by Hashwanth Sutharapu - SDE at MAQ Software (Microsoft Partner), building enterprise agents for MCAPS teams. Contributor to Microsoft Agent Framework (9.3K stars) and awesome-copilot (29K stars). MS CS from Arizona State University (3.93 GPA).

• GitHub: github.com/hashwnath
• LinkedIn: linkedin.com/in/sutharapuhashwanth
• Email: hsuthara@asu.edu