MCP Resources

General definition

MCP Resources are read-only pieces of context that a Model Context Protocol server exposes to a client: files, documentation, database schemas, configuration, API responses, anything the model may need to read but not act on. Each resource is identified by a URI (RFC 3986) and carries a name, an optional title and description, a MIME type and an optional size in bytes. Contents come back as text or as a base64 blob. The specification calls resources application-driven: the host application decides how they are surfaced, whether as a picker the user chooses from, a searchable list or automatic inclusion in context.

The protocol methods are:

• resources/list: enumerate the available resources, paginated
• resources/read: fetch the contents of one URI
• resources/templates/list: expose parameterised resources as RFC 6570 URI templates, for example file:///{path}
• resources/subscribe and notifications/resources/updated: optional change notifications for a single resource
• notifications/resources/list_changed: the catalogue itself changed (servers declare listChanged)

Three primitives, three controllers. Tools are model-controlled and may have side effects. Prompts are user-controlled templates. Resources are application-controlled reading material. A tool result can also return a resource link or embed a resource, which is how a server hands the model a document it did not know to ask for. Custom URI schemes are allowed (a server can define its own myapp://) as long as they follow RFC 3986; https:// should only be used when the client can fetch the URL itself without going through the server.

Why it matters: resources let a server ship documentation, quickstarts and reference material next to its tools, so the assistant can read how to use the server before calling anything, without that material consuming context on every turn the way tool definitions do. Servers must validate URIs and apply access controls to sensitive resources, and binary data must be properly encoded.

In the Ethora ecosystem

The Ethora MCP server publishes four resources under a custom ethora:// scheme: ethora://docs/auth-map (which token type each route accepts and which auth mode to be in), ethora://docs/chat-component/quickstart (the React chat component), ethora://docs/sdk-backend/quickstart (the backend SDK) and ethora://docs/recipes (the step-by-step recipes that the ethora-help tool also returns). They are listed with the tools on the reference page.

The same material is reachable through the search and fetch tools (the ChatGPT connector convention) for clients that do not surface resources at all, and the six prompts wrap the same documents as ready-to-run instructions. Resources cost nothing until read, so an assistant can list them at connect time and read only the one it needs, which keeps the AI SDK docs out of the context window until they are useful.