Universal MCP toolkit + agent runtime to expose tools, connect multiple MCP servers (HTTP/stdio), and orchestrate them with LLMs — with production controls (budgets, logs, retries, redaction).

For who

• Devs building MCP servers (Python/TypeScript) and want “tools in minutes”
• Teams shipping agent workflows across many tools/servers
• Anyone who needs production guardrails (cost, security, observability)

What you get

• ✅ Expose Python functions as MCP tools (HTTP, stdio, in-process)
• ✅ Unified agent to orchestrate tools across multiple servers
• ✅ PolyMCP Inspector: test/debug/monitor MCP servers in a web UI
• ✅ Security + reliability: redaction, allowlists, retries, health checks
• ✅ Skills system: load only relevant tools (token savings)
• ✅ LLM providers: OpenAI, Anthropic (Claude), Ollama, DeepSeek, Kimi (e custom providers)

---

🎉 What's New (open)

PolyMCP PYTHON VERSION UPDATE

Authentication System

• 🔐 Pluggable auth: JWT, OAuth2 (RFC 6749), API keys, custom providers
• 🔄 Auto-refresh with exponential backoff + jitter
• 🛡️ Thread-safe token management & automatic retry on 401/403
• 🎯 Drop-in auth wrapper for UnifiedPolyAgent (no code changes needed)

Security & Isolation

• 🐳 Docker executor: run code in isolated containers

Production Stdio Client

• 🔧 Robust JSON-RPC line parsing with timeout handling
• 🛡️ Graceful shutdown: proper pipe cleanup & process termination
• 🎭 Battle-tested with Playwright, filesystem, and custom servers

Enhanced Security & Reliability

• 🔒 Multi-layered auth with circuit breakers
• 📊 Health checks, structured logs, and metrics
• 🔄 Transient failure retry with intelligent backoff
• 🛡️ Input validation, schema enforcement, workspace isolation

---

Install

pip install polymcp

Quickstart (60s): expose 2 tools via MCP (HTTP)

Create server.py:

from polymcp.polymcp_toolkit import expose_tools_http

def greet(name: str) -> str:
    """Say hello."""
    return f"Hello, {name}!"

def add(a: int, b: int) -> int:
    """Add two numbers."""
    return a + b

app = expose_tools_http(tools=[greet, add], title="My MCP Tools")

Run:

uvicorn server:app --reload

You now have:

• GET /mcp/list_tools
• POST /mcp/invoke/<tool_name>

---

IMPORTANT: MCP base URL

When connecting an agent or generating skills, always use the MCP base path:

✅ Correct: http://localhost:8000/mcp
❌ Wrong: http://localhost:8000

Quick check:

• GET http://localhost:8000/mcp/list_tools should return your tools list
• POST http://localhost:8000/mcp/invoke/<tool_name> invokes a tool

If you see 405 Method Not Allowed on http://localhost:8000/, you're hitting the wrong endpoint (missing /mcp).

Quickstart: run an agent that uses your tools

import asyncio
from polymcp.polyagent import UnifiedPolyAgent, OpenAIProvider

async def main():
    async with UnifiedPolyAgent(
        llm_provider=OpenAIProvider(),
        mcp_servers=["http://localhost:8000/mcp"],
    ) as agent:
        out = await agent.run_async("Greet Luca and then add 5 + 10")
        print(out)

asyncio.run(main())

---

Production config (safe defaults)

from polymcp.polyagent import UnifiedPolyAgent

agent = UnifiedPolyAgent(
    llm_provider=llm,

    # Budget controls
    max_tokens=100000,
    max_tool_calls=20,
    max_wall_time=300.0,

    # Security
    redact_logs=True,
    tool_allowlist={"greet", "add"},

    # Observability
    enable_structured_logs=True,
    log_file="agent.log",

    # Resilience
    max_retries=3,
    enable_health_checks=True,
    enable_rate_limiting=True,

    # Architecture
    use_planner=True,
    use_validator=True,
)

---

PolyMCP Inspector (test/debug/monitor)

polymcp inspector
# opens http://localhost:6274

Inspector highlights

• Multi-server dashboard (HTTP + stdio)
• Interactive tool testing + live request/response
• Metrics (latency, success rate, calls)
• Export reports (json/markdown/html)
• Test suites for regression

---

Common workflows

1) Connect multiple servers (HTTP + stdio)

agent = UnifiedPolyAgent(
    llm_provider=llm,
    mcp_servers=["http://localhost:8000/mcp"],
    stdio_servers=[{"command": "npx", "args": ["@playwright/mcp@latest"]}],
)

2) Stdio MCP server (Claude Desktop compatible)

from polymcp import expose_tools_stdio

def calculate(a: int, b: int) -> int:
    return a + b

server = expose_tools_stdio(
    tools=[calculate],
    server_name="Math Tools",
    server_version="1.0.0"
)

if __name__ == "__main__":
    server.run()

3) Skills system (load only relevant tools)

Generate skills (CLI)

Skills are generated by discovering tools from your MCP servers and auto-categorizing them.

HTTP servers

Use the MCP base URL (include /mcp):

polymcp skills generate --servers "http://localhost:8000/mcp" --output ./mcp_skills --verbose

Stdio servers (Playwright, filesystem, etc.)

Enable stdio discovery:

polymcp skills generate --stdio --servers "npx -y @playwright/mcp@latest" --output ./mcp_skills --verbose

Then enable them in your agent:

agent = UnifiedPolyAgent(
    llm_provider=llm,
    skills_enabled=True,
    skills_dir="./mcp_skills",
    mcp_servers=["http://localhost:8000/mcp"],   # optional if using HTTP tools too
    stdio_servers=[{"command":"npx","args":["-y","@playwright/mcp@latest"]}],  # optional
)

agent = UnifiedPolyAgent(
    llm_provider=llm,
    skills_dir="./skills",
    mcp_servers=["http://localhost:8000/mcp"],
)

---

TypeScript implementation (PolyMCP-TS)

PolyMCP also ships a TypeScript implementation for Node/TS ecosystems:

• schema validation (Zod)
• HTTP + stdio servers
• agent framework + tooling

📖 See: polymcp-ts/README.md

---

Docs & Links

• GitHub: https://github.com/llm-use/polymcp
• PyPI: https://pypi.org/project/polymcp/
• Website: https://www.poly-mcp.com
• Examples: examples/

---

Star History

License

MIT