MCP server

MCPServer wraps a FastMCP app as an ASGI endpoint. Tools defined in Python become callable over the Model Context Protocol from Claude, Cursor, or your own agents, running on the cluster with the image, secrets and GPU you give it.

server.py

from fastmcp import FastMCP
from levers9 import Blueprint
from levers9.integrations import MCPServer

mcp = FastMCP("levers-tools")

@mcp.tool()
def word_count(text: str) -> int:
    return len(text.split())

server = MCPServer(
    mcp,
    name="tools",
    cpu=1, memory="512Mi",
    image=Blueprint(python_version="python3.11", python_packages=["fastmcp"]),
    secrets=["OPENAI_API_KEY"],
    authorized=True,                   # clients send the workspace token
)

levers9 serve server.py:server         # a temporary URL while you build
levers9 deploy server.py:server --name tools
claude mcp add --transport http tools <the URL>/mcp \
  --header "Authorization: Bearer <workspace token>"

levers9 deploy prints the URL; FastMCP’s streamable HTTP transport answers at /mcp under it, and clients send the workspace token as a bearer header. A JSON client config carries the same header:

{ "mcpServers": { "tools": { "url": "<the URL>/mcp",
                             "headers": { "Authorization": "Bearer <workspace token>" } } } }

Pass authorized=True every time. The server holds your secrets and maybe a GPU, and a public one lets anyone with the URL call every tool. authorized=False makes it public on purpose, for example for a client that can’t send a header. If you leave it out, the server is public, as it always was, and deploy and serve warn you; a later release makes True the default. MCPServerArgs(fastmcp_version="2.7.1") pins the FastMCP release installed in the image. Everything from Parameters applies: keep_warm_seconds, autoscaler, volumes, gpu.

Get started

Build with the SDK

Dashboard

Data Lab

RL Arena

Operate