MCP Server Development

Time to BUILD: a production MCP server — typed tools, exposed resources, tests, an Inspector, deployment with auth, and observability. FastMCP makes the first 80% minutes, not weeks.

▶ Watch this reel

What you'll learn

FastMCP: first server
Resources & typed design
Testing with the Inspector
Deployment: HTTP, auth, observability

Remember this

FastMCP: typed functions + docstrings become tools/resources in minutes — your signature is the schema, your docstring is the model's documentation
Design small surfaces: resources for data, prompts for repeated shapes, explicit handles for state; test with the Inspector plus scripted protocol tests
Remote deployment = operating a service: streamable HTTP, spec-aligned OAuth, per-call observability, pinned revisions and deprecation tracking

FastMCP

Decorators: typed function → tool; docstring → description; Pydantic → schema.
C# SDK mirrors it (attributes + ASP.NET hosting).
You still own AG-02: descriptions, validation, actionable errors.

Design

Resources with mental-model URIs · prompts for repeated shapes · structured outputs.
Small surfaces: ~5 legible tools; split by domain.
Stateless: explicit handles as arguments for multi-call state.

Testing

MCP Inspector: interactive poke-list-invoke loop — test the model's experience.
Scripted: in-process sessions in pytest — schema/docs drift, failure quality, round-trips.

Deployment

streamable HTTP · load balancer friendly · spec-aligned OAuth (or gateway).
Observability: per-call structured logs (hash args, log user), RED metrics, cost/call.
Pin spec revision · track deprecations · backward-compatible surface.

Code: The same server, deployed: HTTP + logging + auth hook

from mcp.server.fastmcp import FastMCP
import logging, structlog

log = structlog.get_logger("mcp.orders")
mcp = FastMCP("orders")

@mcp.tool()
def search_orders(customer: str, limit: int = 5) -> list[dict]:
    """Search orders by customer name fragment … (same docstring!)"""
    log.info("tool.call", tool="search_orders",
             arg_hash=hash((customer, limit)),      # not raw args
             user=current_user_id())               # auth context from hook
    ...

# Local:   mcp.run()                          # stdio
# Remote:  mcp.run(transport="http",
#                  host="0.0.0.0", port=8080,
#                  auth=oauth_provider)       # spec-aligned OAuth
#
# Ops:  per-call logs · latency histograms · error-rate alerts ·
#       cost-per-tool dashboards · pinned spec revision in CI