FastMCP (standalone)

Warnings

• breaking v3.0 (February 2026) is a major breaking rewrite from v2.x. Anyone not pinning fastmcp was auto-upgraded to v3.0 on pip install. Major changes: provider architecture replaces direct component registration, ctx.set_state()/ctx.get_state() are now async, StreamableHttpTransport replaces older HTTP transport classes, auth providers no longer auto-load config.
  Fix: Pin to fastmcp<3 for v2 stability while migrating. See gofastmcp.com/development/upgrade-guide for the v2→v3 upgrade guide and copyable LLM migration prompt.
• breaking In v3, @mcp.tool decorator returns the original function unchanged (so decorated functions remain directly callable). In v2.x, @mcp.tool returned a FunctionTool object. Code accessing .name, .description, or other FunctionTool attributes on the decorated result will break.
  Fix: Remove any attribute access on decorated functions. Set FASTMCP_DECORATOR_MODE=object temporarily for compatibility (deprecated).
• breaking Transport config (host, port) must be passed to mcp.run(), not the FastMCP() constructor. Passing them to the constructor raises TypeError.
  Fix: Replace FastMCP('server', host='0.0.0.0', port=8080) with FastMCP('server') and mcp.run(transport='http', host='0.0.0.0', port=8080).
• breaking fastmcp.server.auth module (introduced v2.12.0) is explicitly exempt from semantic versioning stability guarantees and has breaking changes on minor versions. Auth API changed significantly in v3.
  Fix: Pin to an exact fastmcp version if using auth in production. Do not rely on auth API stability across minor versions.
• breaking GitHub repository moved from jlowin/fastmcp to PrefectHQ/fastmcp with v3.0. Old GitHub links forward automatically. PyPI package name, import paths, and CLI commands are unchanged.
  Fix: Update bookmarks and CI checkout references from jlowin/fastmcp to PrefectHQ/fastmcp.
• gotcha fastmcp includes the mcp package as a dependency — installing both separately can cause version conflicts if mcp is pinned independently. fastmcp pins its compatible mcp version internally.
  Fix: If using fastmcp, do not independently pin mcp in requirements. Let fastmcp manage the mcp dependency version.
• gotcha Minor versions may contain breaking changes (explicitly documented in release policy). 'Semantic versioning is pragmatic' — do not assume minor bumps are safe without checking release notes.
  Fix: Pin to exact version in production: fastmcp==3.0.2. For development use ~=3.0 to get patches only.
• gotcha fastmcp and the mcp SDK's bundled FastMCP 1.0 (from mcp.server.fastmcp) have the same class name but different behavior. Mixing imports in one project causes silent behavioral differences.
  Fix: Use only one: either 'from fastmcp import FastMCP' (standalone) or 'from mcp.server.fastmcp import FastMCP' (bundled v1). Never mix both in one project.
• gotcha CVE-2026-24486 (python-multipart) and CVE-2026-0994 (protobuf) were patched in fastmcp 3.0.2. Earlier 3.x and 2.x versions are vulnerable.
  Fix: Upgrade to fastmcp>=3.0.2.

Install

• pip install fastmcp Python (latest — v3.x)
• pip install 'fastmcp<3' Python (pin to v2.x — use if upgrading from v2 not yet complete)
• pip install 'fastmcp[openai]' Python with OpenAI client integration
• pip install 'fastmcp[anthropic]' Python with Anthropic client integration

Imports

• FastMCP (v2 and v3 — same import path)

  from fastmcp import FastMCP
  
  mcp = FastMCP('My Server')
  
  @mcp.tool
  def add(a: int, b: int) -> int:
      """Add two numbers"""
      return a + b
  
  mcp.run()

  Use 'from fastmcp import FastMCP' for standalone package. 'from mcp.server.fastmcp import FastMCP' is FastMCP 1.0 bundled in the official mcp package — a different, older version.
• Client

  from fastmcp import Client
  
  async with Client('http://localhost:8000/mcp') as client:
      tools = await client.list_tools()
      result = await client.call_tool('add', {'a': 1, 'b': 2})

  Client class replaces any earlier transport-specific client patterns. Handles transport negotiation automatically.

Quickstart

Decorators derive tool name, description, and JSON schema from function signature and docstring automatically. mcp.run() defaults to stdio for local use; pass transport='http' for remote deployments.

# Server
from fastmcp import FastMCP

mcp = FastMCP('Demo')

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

@mcp.resource('data://{name}')
def get_data(name: str) -> str:
    """Fetch named data"""
    return f'Data for {name}'

@mcp.prompt
def review(code: str) -> str:
    return f'Review this code:\n{code}'

if __name__ == '__main__':
    mcp.run()  # stdio by default
    # mcp.run(transport='http', host='0.0.0.0', port=8000)  # remote

# ---

# Client
import asyncio
from fastmcp import Client

async def main():
    async with Client('http://localhost:8000/mcp') as client:
        tools = await client.list_tools()
        result = await client.call_tool('add', {'a': 1, 'b': 2})
        print(result)

asyncio.run(main())

# ---

# CLI usage
# fastmcp list http://localhost:8000/mcp     # list tools on any server
# fastmcp call http://localhost:8000/mcp add a=1 b=2  # call a tool
# fastmcp discover                            # find all locally configured servers