Building an MCP server

Anthropic FDE and agent-engineer interviews literally ask candidates to build an MCP server. The shape is small: declare a server, register tools/resources/prompts, and run a transport. Here is the essence with the Python SDK.

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("orders")

@mcp.tool()
def get_order_status(order_id: str) -> str:
    """Look up the status of a customer order by id."""
    return lookup(order_id)  # your real logic

@mcp.resource("orders://recent")
def recent_orders() -> str:
    """Recent orders as context."""
    return format_recent()

if __name__ == "__main__":
    mcp.run(transport="stdio")   # or "streamable-http" for remote

• The SDK turns typed function signatures + docstrings into the JSON schema + description the model sees — so name and document them well.
• Test locally with the MCP Inspector, then point a client (Claude Desktop/Code, your agent) at it.
• For production: pick streamable HTTP, add auth, apply least-privilege, and log every call.
• Don't load every tool into context: tool definitions at scale blow the attention budget — use progressive disclosure / tool search, loading definitions on demand.
• Code mode: for tool-heavy work, let the model write code that calls MCP tools in a sandbox, so only final results enter context — order-of-magnitude token savings.
• Tool poisoning: malicious tool names/descriptions/results are an injection vector — vet and pin servers, and review tool metadata like third-party code.