An MCP tool is an executable capability supplied through the Model Context Protocol. Suppliers define tools; AI applications and LLM agents consume them to extend their capabilities. section 5.1.1
An MCP client sends a tool-execution request to an MCP server, the external system that supplies the tool. Multiple tools can share a server, each with a corresponding server-side definition. Servers can run locally over stdio or remotely over streamable HTTP. section 5.1.3
Figure 5.4 The MCP architecture integrated into LLM agents and our framework

See how MCP servers, clients, and tools connect to an LLM agent and its framework. section 5.1.3
A framework can expose MCP tools to an agent through MCPTool and MCPToolProvider. In a multi-agent system, a tool can be assigned to one agent or selected agents; shared providers must also handle concurrent connection setup. section 5.0 section 5.4
