What is an MCP resource?

MCP resource, explained#

Tools answer "do something". Resources answer "here is something to read". The specification describes resources as application-driven: the host decides how to surface them, for instance as a picker where a person attaches a document to the conversation, or by including some automatically.

Each resource has a uri, a name, and optionally a title, description and mimeType. The URI can use any scheme the server chooses, such as file:// or a custom one like app://customer/42. Reading it returns one or more contents, each either text or base64 blob data.

Servers can also publish resource templates, URI patterns with variables in RFC 6570 style such as app://project/{slug}. A client can offer completion for the variable and then read the resulting URI. Servers may let clients subscribe to a resource and notify them when it changes, and may notify when the list itself changes.

Resources and tools often sit side by side. A server might expose a read-only resource for a record and a separate tool to change it, so attaching context never risks a write, and the same permission checks apply to both.

Resources are useful when the context is the same no matter what the model is about to do, and when you want the person, not the model, to decide what goes in. They also cost the model no tool call: the content is simply there.

For example, a project tool exposes app://project/{slug}. Before asking for a status report, a manager attaches app://project/checkout-redesign. The assistant now has the project's summary, open work and recent updates in context, and can write the report without first searching for them.