How Tool, Function Calling, MCP, and Skill Work Together
A tutorial uses a synthetic document-comparison scenario to show how Tool, Function Calling, MCP, and Skill participate in one assistant workflow. It describes tools as executable capabilities, function calling as a structured way for a model to request a tool, MCP as a connection to external services and their tools, and a Skill as reusable task guidance and resources. The runnable example starts a standard-input/output MCP process to retrieve synthetic pages, then saves a report. It also shows how to inspect call parameters, paired results, action records, and generated artifacts. Because the demo uses preset model actions, it does not establish whether a model can independently select the correct Skill or tool.
The tutorial provides a component-by-component troubleshooting framework: method problems can be checked through Skill loading, capability problems through tool calls, connection problems through MCP discovery and execution, and result problems through artifacts and receipts. This can help users identify which layer needs adjustment instead of changing every component at once.
Tool、Function Calling、MCP、Skill，一次行动怎样发生？
你准备给助手增加“读取产品指南”的能力。搜索一圈，会同时看到 Tool、Function Calling、MCP、Skill。每个介绍都能举出读取文档的例子，究竟应该从哪里开始？
可以先跟踪一次具体行动。用户要求比较 Atlas Notes 两版指南，助手需要取得 A 版第一页。这是我们公开示例中的合成资料。先看取得这一页需要哪些东西，再把名称放回对应位置。
先有一项实际能力
应用需要一段真正能够取得页面的程序。教学上，可以写成一个普通函数。
async function readPage({ documentId, page }) {
return documentStore.readPage(documentId, page)
}
#Model Context Protocol (MCP)#Agent Skills#Anthropic#Open ETF Agent