mcpzero (Python)

Python SDK for MCPZERO (open source, MIT). Expose your own MCP server through the MCPZERO tunnel — a public Streamable-HTTP endpoint plus full dashboard visibility — directly from code, without the CLI.

pip install mcpzero-sdk          # embedded proxy only
pip install "mcpzero[mcp]"   # + in-process adapter for the official MCP SDK

Create an endpoint and a management key in the dashboard. A management key is a user-level credential that lets the SDK publish (register a tunnel for) any endpoint you own — the headless alternative to mcpzero login. By convention they are read from MCPZERO_ENDPOINT_ID / MCPZERO_MGMT_KEY (gateway base from MCPZERO_GW_BASE, default https://gw.mcpzero.io).

Not to be confused with an API key, which is the consumer-side credential AI clients use to call your published endpoint.

Tunnel an in-process MCP server (primary API)

Run your official mcp SDK server with its streams backed by the tunnel:

import asyncio
import mcpzero
from mcp.server.lowlevel import Server

server = Server("my-mcp")
# ... @server.list_tools() / @server.call_tool() ...

asyncio.run(mcpzero.serve(server, endpoint_id="ep_...", management_key="mzm_..."))

Or drive the streams yourself:

async with mcpzero.tunnel(endpoint_id="ep_...", management_key="mzm_...") as (read, write):
    await server.run(read, write, server.create_initialization_options())

Tunnel a local MCP server (lower-level primitive)

A library version of mcpzero tunnel, for a server running as a separate process or HTTP service:

# stdio:
handle = await mcpzero.tunnel_proxy(command=["python", "server.py"])
# or HTTP:
handle = await mcpzero.tunnel_proxy(url="http://localhost:3000/mcp")

print("Live at", handle.endpoint_url)
# ... later:
await handle.aclose()

Visibility

In tunnel mode, visibility is produced by the gateway automatically — the dashboard reflects calls exactly as it does for the CLI tunnel. No extra code is needed.

See ../PROTOCOL.md for the tunnel wire protocol.

Docs: https://mcpzero.io/docs/sdk/tunnel/python