Skip to main content
Claude Code supports the Model Context Protocol (MCP), which allows external servers to expose tools and resources that Claude can use. Three built-in tools handle MCP interaction: the MCP tool (for calling server-defined tools), ListMcpResources (for discovering available resources), and ReadMcpResource (for reading a specific resource by URI).

​
MCP (server-defined tools)

When an MCP server is connected, each tool it exposes becomes available to Claude as a first-class tool with its own name, description, and input schema — all defined by the server. Claude invokes these tools the same way it invokes built-in tools.

​
How it works

Claude Code uses the MCP tool mechanism as a base, but each connected server’s tools are registered with:
  • A real tool name (e.g., github__search_repositories, postgres__query)
  • A description provided by the server
  • An input schema defined by the server
The input parameters are entirely determined by the server’s tool definition. There is no fixed parameter set for MCP tools — each tool accepts whatever the server specifies.

​
Output

​
result
string
The MCP tool’s execution result, returned as a string. The content and format depend entirely on the server and tool.

​
Referencing MCP tools

You can ask Claude to use a specific MCP tool by name: 
Use the github__search_repositories tool to find TypeScript projects with over 1000 stars.
Or refer to it by its MCP server: 
Search for recent commits using the GitHub MCP server.

​
Permissions

MCP tools use a passthrough permission model. Claude Code will ask for your consent before invoking an MCP tool for the first time. You can pre-approve MCP tools in your settings.
If an MCP server provides a web fetch or search tool, prefer it over the built-in WebFetch or WebSearch tools — MCP tools may have authenticated access or fewer restrictions.

​
ListMcpResources

Lists the resources available from connected MCP servers. Resources are data artifacts exposed by a server (files, database records, API responses, etc.) that can be read by URI.

​
Parameters

​
server
string
Optional. Filter results to a specific MCP server by name. When omitted, resources from all connected servers are returned.

​
Output

Returns an array of resource objects:
​
uri
string
The resource URI — used as the identifier to read the resource with ReadMcpResource.
​
name
string
Human-readable resource name.
​
server
string
The name of the MCP server that provides this resource.
​
mimeType
string
MIME type of the resource content (e.g., "text/plain", "application/json").
​
description
string
Optional description of what the resource contains.

​
Examples

List all resources from all servers
{}
List resources from a specific server
server: "myserver"

​
ReadMcpResource

Reads the content of a specific resource from an MCP server, identified by its server name and URI.

​
Parameters

​
server
string
required
The name of the MCP server to read the resource from. Must match a connected server’s name exactly.
​
uri
string
required
The URI of the resource to read. Obtain URIs from ListMcpResources.

​
Output

​
contents
array
Array of content objects. Each object includes:

​
Usage notes

  • Binary blob content (images, PDFs, etc.) is automatically saved to a temporary file rather than inlined as base64. The blobSavedTo field gives you the path; use the Read tool to access it.
  • The server must have resource capabilities enabled. If it doesn’t, ReadMcpResource will return an error.

​
Example workflow

Step 1: Discover available resources
ListMcpResources({ server: "my-db-server" })

Step 2: Read a specific resource
ReadMcpResource({
  server: "my-db-server",
  uri: "db://schema/users"
})

​
Connecting MCP servers

MCP servers are configured in your Claude Code settings. See the MCP configuration guide for instructions on adding servers. Once connected, Claude Code automatically discovers available tools and resources from each server at startup and when servers send change notifications.