Content Blocks and Structured Output

The question this answers

Every result so far had content be a one-item list: [TextContent(text="5")]. It would be reasonable to wonder why it's a list at all, instead of just a string. This lesson is the answer: a single tool call can return more than one kind of content at once.

Returning more than text

FastMCP recognizes a few return shapes beyond a plain string. To return an image, use the Image helper.

from mcp.server.fastmcp import FastMCP, Image

@mcp.tool()
def make_thumbnail() -> Image:
    """Return a small PNG thumbnail."""
    png_bytes = render_thumbnail()  # however you produce the bytes
    return Image(data=png_bytes, format="png")

The client receives an ImageContent block instead of TextContent, with .mimeType ("image/png") and .data (base64-encoded bytes) instead of .text.

Returning several blocks at once

Return a list, and each item becomes its own content block, mixed types allowed.

@mcp.tool()
def make_thumbnail_with_caption() -> list:
    """Return a caption and a thumbnail together."""
    return ["Here is your thumbnail:", Image(data=png_bytes, format="png")]

result.content
# [TextContent(text="Here is your thumbnail:"), ImageContent(mimeType="image/png", data="...")]

This is the actual reason content is a list everywhere in this course, one tool call, one API response, but potentially several distinct pieces of content, exactly like a chat message from a multimodal model can contain both text and an image in the same turn.

structuredContent: the typed escape hatch

Lesson 12 mentioned structuredContent briefly. When a tool's return type is a simple annotated type (int, a TypedDict, a Pydantic model), FastMCP includes a second, machine-readable version of the result alongside the human-readable content.

result.content            # [TextContent(text="5")]
result.structuredContent  # {"result": 5}

content is what you show a human or feed back to a model as conversation text. structuredContent is what your own code should read when it needs the actual typed value, no string-parsing required.