MCP Tools, Resources & Prompts: Core Capabilities Explained
A practical map of the capabilities an MCP server can expose and a client can consume, updated for the 2026-07-28 specification.
What are MCP core capabilities?
Model Context Protocol gives clients and servers a common way to describe capabilities such as tools, resources, and prompts. Tools represent actions, resources expose retrievable context, and prompts provide reusable prompt templates. The protocol also defines supporting behavior such as schemas, elicitation, completion, pagination, progress, cancellation and caching.
If you are learning agents first, start with the AI Agents tutorial for developers. For the full protocol sequence, use the complete MCP course.
Actions a model or client can request from a server.
Names, boundaries, confirmation and predictable behavior.
Input/output contracts and JSON Schema design.
Expose retrievable application or domain context.
Address and parameterize retrievable resources.
Reusable prompt templates surfaced by servers.
Understand where each type of instruction belongs.
Request additional information during an interaction.
MCP tools vs resources vs prompts
| Capability | Best mental model | Example |
|---|---|---|
| Tools | Do something | Create a ticket, query a database, send a request |
| Resources | Read something | Repository file, customer record, schema, documentation |
| Prompts | Reuse an interaction template | Code review prompt, incident analysis prompt |
Key 2026 changes
The current protocol makes list/read results cacheable, adopts full JSON Schema 2020-12 for tool input/output schemas, restructures server-to-client interactions around Multi Round-Trip Requests, and deprecates roots, sampling and logging for new implementations. Deprecated features may still exist during their policy window, but new implementations should follow current replacements and specification guidance.
Where to go next
- Start with MCP tools if you are building actions.
- Read MCP resources if your server exposes context.
- Read MCP prompts if you need reusable prompt templates.
- Return to the AI Agents learning path to see how MCP fits inside an agent architecture.