Resources

Tools do things, resources are things

Every tool so far does something: converts a number, adds two values. A resource is different: it just hands back data for context, nothing is "done," nothing changes. Think of it like reading a file versus running a script.

@mcp.resource("notes://today")
def today_note() -> str:
    """Today's note."""
    return "Buy milk."

@mcp.resource(uri) takes a URI instead of relying on the function name. notes://today is not a real network address, it's just a unique identifier the client uses to ask for this specific piece of data later.

Discovering and reading a resource

The same list/get pattern as tools, with different verbs:

• resources/list (here: await mcp.list_resources()): what resources exist?
• resources/read (here: await mcp.read_resource(uri)): give me the content at this URI.

resources = await mcp.list_resources()
# [Resource(uri='notes://today', name='today_note', description="Today's note.", ...)]

content = await mcp.read_resource("notes://today")
# [ReadResourceContents(content='Buy milk.', mime_type='text/plain')]

URI templates: parameterized resources

A resource's URI can include a placeholder, letting one function serve many different resources:

@mcp.resource("notes://{date}")
def note_for_date(date: str) -> str:
    """Get the note for a given date."""
    return f"Note for {date}: nothing yet."

Reading notes://2026-08-03 calls this function with date="2026-08-03", the same way a URL path parameter works in a web framework. This is how a server can expose an entire, dynamic dataset (every note you've ever written) as one resource definition instead of one per item.

When to use a resource instead of a tool

If an AI needs to decide to fetch something based on the conversation, a tool works fine, that's the whole point of tool calling. Resources exist for the other case: data a host wants to pull in as context up front, or attach to a message, without the AI needing to explicitly ask for it via a function call. The distinction matters more to host applications, how they present context to a user, than to the AI itself, but the primitive exists so a server can be explicit about which of its data is "context to attach" versus "action to decide on."