What is an MCP tool?

MCP tool, explained#

Tools are how a model acts rather than only talks. The specification calls them model-controlled: the language model reads the tool descriptions, decides that one fits the request, and fills in its arguments. The host is expected to keep a person in the loop, showing which tool is about to run and letting them approve or refuse it.

A tool definition is small. It has a unique name, a human-readable title and description, an inputSchema in JSON Schema that says which arguments exist and which are required, and optionally an outputSchema describing structured results. It can also carry annotations such as readOnlyHint or destructiveHint. Those are hints for the host's interface, and the specification warns that hints from an untrusted server should not be relied on.

A tool result is a list of content items, usually text, and may include structuredContent matching the output schema. When a tool fails in a way the model should know about, such as invalid input or a refused action, the server returns a normal result with isError: true and a message, rather than a protocol error. That way the model can read what went wrong and try something else.

Good tools are narrow and well described. A model chooses tools from their descriptions, so a vague description leads to wrong calls, and one tool that does ten things is harder to use safely than three that each do one.

For example, a tracker's create_issue tool might take title (required), project and priority. Asked to "log a bug that the export button is broken", the model calls it with a title and priority, gets back the new issue's id and link, and tells the person it is done.