MCP in a Minute: Tools and resources

In one sentence

MCP tools perform actions; MCP resources expose readable context.

Why it matters

If you model everything as a tool, agents get noisy action menus. If you hide actionable operations behind “resources,” agents cannot do the job. The distinction keeps interfaces honest.

How it works

• Tool — something to do: create an issue, run a query, trigger a deploy preview.
• Resource — something to read: a file-like document, a schema, a transcript, a config snapshot.

Clients discover what a server offers, then the model selects tools during reasoning and may pull resources into context when needed.

Design tools with clear names, tight inputs, and predictable side effects. Design resources to be fetchable without surprise mutation.

Example

For a CI system:

• resource: latest failing job log;
• tool: rerun workflow on branch X.

That split lets an agent inspect first and act second.

Agentic coding use

Coding agents already edit files and run shells. MCP should extend that with structured capabilities you would not want expressed as ad-hoc bash. Prefer small tools with verification over mega-tools that “do the release.”

Watch out

Tool results become context. Verbose tools can flood the window and degrade later reasoning. Return summaries plus pointers when payloads are large.