FastMCP is a Python framework for building and connecting MCP tools. Testing the data contract before adding an AI host makes failures easier to locate. This 4.2.0 test used an in-memory client to check discovery, a known response, and an error. It opened no HTTP port and did not ask a model to choose or call a tool.
After starting the server, check its interface through MCP Inspector's tool calls. For file-tool scope, refer to Filesystem MCP's allowed directories.
Define a tool with a known answer
The dispatch_cutoff tool accepts a string note. NOTE-31 returns 08:30; any other identifier raises ValueError. Keep the first tool small so a data-contract failure is visible before adding authentication, a database, or a model.
Pin and exercise the local contract
Run the shell block below in a Python 3.12 virtual environment to install the pinned package and execute the embedded Python example. Client(mcp) connects directly to the server object. That exercises the in-memory transport and does not establish that stdio, HTTP, or a remote host works.
These commands use a POSIX shell. The test ran in Linux; Windows shell behavior was not tested.
python -m pip install "fastmcp==4.2.0"
python - <<'PY'
import asyncio
from fastmcp import FastMCP, Client
mcp = FastMCP("DispatchFixture")
@mcp.tool
def dispatch_cutoff(note: str) -> str:
if note != "NOTE-31":
raise ValueError("Unknown synthetic dispatch")
return "08:30"
async def check():
async with Client(mcp) as client:
print(await client.list_tools())
print(await client.call_tool("dispatch_cutoff", {"note": "NOTE-31"}))
asyncio.run(check())
PY
Three observed checks
list_tools exposed dispatch_cutoff with a required string note. The valid call returned 08:30 in text and a structured result. An unknown identifier raised ToolError with the synthetic failure reason. These checks validate the interface; correct tool selection by a model needs a separate test.
Evidence still needed before deployment
HTTP, stdio, OAuth, authorization, and Gateway behavior are described from official v4.2.0 material and the saved upstream snapshot; they were not locally tested. An in-memory success does not make external inputs trustworthy or provide operating-system isolation. Transport, timeouts, permissions, and real data sources need their own checks before connecting an AI host.
Official sources and test scope
Codex ran this synthetic-data test in a credential-free, non-root, network-disabled container. It is not a claim that the site owner tested it. Package version: 4.2.0; documentation snapshot: a8745d61e650.