Tool use is the broader concept of which function calling is one implementation. Anthropic uses "tool use" as the formal name in the Claude API; OpenAI calls the same thing "function calling" or "tools"; Google calls it "function calling". The semantics are the same: the model decides when to invoke an external action, formats the request as structured data, and the application executes it.
In Anthropic's API, tools are defined in a tools array on the request: each tool has a name, a description (used by the model to decide when to use it), and an input_schema (JSON Schema). The model returns a tool_use block with the chosen tool and arguments. The application returns a tool_result in the next request. Multi-turn tool sequences are common: the model might call get_weather, get the result, then call send_email based on what it saw.
Best practice is to write tool descriptions for the model (not for humans) and to keep tools narrow. A vague tool gets used wrong; a specific tool with clear examples gets used well.
Every useful agent uses tools. The quality of the tool definitions determines the quality of the agent's behavior. This is why "tool design" has become a real discipline distinct from API design.
The TerminalFeed MCP server exposes our entire API as well-described tools an AI agent can call. The /api/llm-tools endpoint serves the same definitions in OpenAI and Anthropic formats for direct copy-paste.