MCP Tools vs Resources vs Prompts: Which Should You Expose?

By AHMAD ·

Understand the three MCP server primitives and choose the right one for actions, data, and reusable prompts.

MCP Tools vs Resources vs Prompts: Which Should You Expose?

An MCP server can expose three kinds of capability, and choosing the right one makes an integration easier to understand. The MCP Python SDK (https://py.sdk.modelcontextprotocol.io/get-started/first-steps/) describes tools as operations the model may call, resources as data the application may load, and prompts as templates a user chooses.

A decision rule

Ask who should decide that the capability runs. If the assistant needs to decide whether to search issues, use a tool. If the host loads a known document into context, use a resource. If a person chooses a named workflow such as “review this pull request,” a prompt fits. The same backend API might support more than one of these, but the user experience is different.

A tool should have a narrow verb and parameters: search_issues(query, limit) is easier to review than run_any_request(url, body). A resource should have a stable identifier and return data without surprising side effects. A prompt should make its purpose and required inputs clear.

Example: a learning library

For a learning site, find_posts(topic) is a tool. learning://post/ is a resource containing one post. summarize_for_beginner could be a prompt template. Publishing a post is a separate write tool with tighter permission and approval than search.

Check before shipping

List every exposed primitive, identify its caller, verify the required permission, and test what happens when a requested item is missing. If a read operation changes data or a prompt silently triggers a write, the interface is misleading. See our introductory MCP guide (https://ubits.tech/learning/p/what-is-mcp-practical-guide-ai-tools-2026) for the host/client/server model.

Primary source: MCP Python SDK first steps (https://py.sdk.modelcontextprotocol.io/get-started/first-steps/).

Editorial note: Prepared with AI assistance and checked against the linked primary documentation on 15 November 2026.