Anthropic Tool Use

Anthropic calls its function-calling feature tool use. You pass a tools array to the Messages API, where each tool has a name, a description, and an input_schema in JSON Schema. When Claude decides to act, the response contains a tool_use content block with a unique ID, the tool name, and an input object — already parsed, not a string — and the response’s stop_reason is set to tool_use. You run the tool and send back a user message containing a tool_result block referencing that ID.

The design difference worth internalizing is that Claude’s responses are lists of typed content blocks, so a single reply can interleave visible text, the model’s reasoning, and one or more tool calls. Anthropic’s docs also lean hard on description quality — detailed, multi-sentence tool descriptions measurably improve Claude’s tool selection — and the same interface underpins Anthropic’s agent products and server-side tools like web search and code execution, so it scales from a one-off script to a full agent.

In practice you’ll define tools in the anthropic SDK, check stop_reason after each call, execute the matching function, and append the tool_result (including is_error: true for failures) before calling again. You’ll use tool_choice to force or disable tools, handle several tool_use blocks in one turn, and reuse the identical schema format when exposing tools over MCP or the Agent SDK.