LLM Agent 工具调用简史：从 Function Call、MCP 到 CLI+Skills（一）
Merged summary
TL;DR - A developer-oriented analysis traces agent tool use from function calling to MCP and newer CLI-plus-Skills workflows. It argues that MCP standardizes portable tool integration rather than creating tool-calling itself, while CLI approaches can be simpler and more token-efficient for local development.
- Function calling lets a model select a tool and arguments; MCP standardizes tool discovery, invocation, and result exchange across AI hosts.
- MCP uses a Host–Client–Server architecture, with tool descriptions and schemas guiding model selection.
- MCP enables reusable integrations across platforms but can introduce configuration, security, context-window, and maintenance overhead.
- CLI tools leverage models’ existing shell knowledge and composability, while MCP remains useful for remote services, structured schemas, cross-platform reuse, and long-running connections.
Sources (1)
LLM Agent 工具调用简史：从 Function Call、MCP 到 CLI+Skills（一）
TL;DR - A developer-oriented analysis traces agent tool use from function calling to MCP and newer CLI-plus-Skills workflows. It argues that MCP standardizes portable tool integration rather than creating tool-calling itself, while CLI approaches can be simpler and more token-efficient for local development.
- Function calling lets a model select a tool and arguments; MCP standardizes tool discovery, invocation, and result exchange across AI hosts.
- MCP uses a Host–Client–Server architecture, with tool descriptions and schemas guiding model selection.
- MCP enables reusable integrations across platforms but can introduce configuration, security, context-window, and maintenance overhead.
- CLI tools leverage models’ existing shell knowledge and composability, while MCP remains useful for remote services, structured schemas, cross-platform reuse, and long-running connections.