What is an MCP server?

What does an MCP server actually expose?

Three kinds of thing, and knowing them makes server descriptions readable:

• Tools: actions the AI can invoke, with typed parameters. “Create an issue”, “run this query”, “post a message”.
• Resources: the data the AI can read, whether files, records or documents.
• Prompts: reusable templates the server offers for common tasks against it.

When do you need one instead of a native integration?

When you want structured access to a tool outside the native set.

Strawberry ships 93 native integrations carrying 1,257 hand-written operations, covering Gmail, Slack, Notion, Salesforce, Linear and Shopify among them, and for those a companion already has typed tools.

Everything else has two routes that both work: drive the site in your logged-in session, or connect it over MCP where a server exists. Internal systems are the clearest MCP case, because your own service can expose exactly the three operations you want a companion to have. The walkthrough is at the MCP integration guide.

What to check before you connect one

The questions are the ones you would ask about any integration.

Who wrote the server. What scopes it requests, and whether it needs write access or only read. Where it runs. What it logs.

One risk is specific to this shape: tool descriptions and returned data are text an AI reads, which makes them a channel for prompt injection. That risk category is open across the industry rather than solved by any one vendor. Prefer narrow scopes, keep approval gates on consequential actions, and read is it safe to let AI use my browser for the wider picture.

Frequently asked questions