Resources, Tools & Prompts

MCP servers expose three kinds of capabilities. Choosing the right one makes your server intuitive.

Tools — actions

The model calls a tool to do something. Every call is a request/response.

@mcp.tool()
def create_issue(title: str, body: str) -> str:
    """Create a GitHub issue."""
    ...

Use tools when the model should change state: write a file, run a query, post a message.

Resources — data

Resources are readable data the model can pull into context, addressed by URI.

@mcp.resource("db://schema")
def schema() -> str:
    """The current database schema."""
    return get_schema()

Use resources for things the model should read on demand: a schema, a style guide, a log file.

Prompts — templates

Servers can ship reusable prompt templates with arguments.

@mcp.prompt()
def review(code: str) -> str:
    return f"Review this code for bugs:\n\n{code}"

Use prompts for repeatable workflows your team standardizes (code review, commit messages).

How the model sees it

From the model's perspective:

• Tools = "things I can call."
• Resources = "things I can read."
• Prompts = "templates I can expand."

A good server uses the right primitive:

| Want the model to… | Use | | --- | --- | | Query the database | tool | | Read the DB schema | resource | | Follow your team's PR checklist | prompt | | Send a Slack message | tool |

Rule of thumb: state changes → tools; facts → resources; recipes → prompts.